Propagation Modelling for Urban Source …...experience which always guide me towards a right direction, and o ering me an open minded research environment. His dedication and commitment

University of Bath

PHD

Propagation Modelling for Urban Source Localization and Navigation

Dai, Zhuangzhuang

Award date:2019

Awarding institution:University of Bath

Link to publication

General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal ?

Take down policyIf you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediatelyand investigate your claim.

Download date: 25. Apr. 2020

https://researchportal.bath.ac.uk/en/studentthesis/propagation-modelling-for-urban-source-localization-and-navigation(6214a3d1-6711-4f13-be07-751ed38b29c2).html

Citation for published version:Dai, Z 2018, 'Propagation Modelling for Urban Source Localization and Navigation', Ph.D., University of Bath.

Publication date:2018

Document VersionPublisher's PDF, also known as Version of record

Link to publication

Attention is drawn to the fact that copyright of this thesis rests with its author. This copy of the thesis has beensupplied on the condition that anyone who consults it is understood to recognise that its copyright rests with itsauthor and that no quotation from the thesis and no information derived from it may be published without theprior written consent of the author. This thesis may be made available for consultation within the UniversityLibrary and may be photocopied or lent to other libraries for the purposes of consultation.

University of Bath

General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

Take down policyIf you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediatelyand investigate your claim.

Download date: 07. Jan. 2019

https://researchportal.bath.ac.uk/en/publications/propagation-modelling-for-urban-source-localization-and-navigation(c749a854-84d3-4635-8925-087e72919eca).html

Propagation Modelling for Urban

Source Localization and

Navigationsubmitted by

Zhuangzhuang Dai

for the degree of Doctor of Philosophy

of the

University of Bath

Department of Electrical and Electronic Engineering

September 2017

COPYRIGHT

Attention is drawn to the fact that copyright of this thesis rests with its author. This

copy of the thesis has been supplied on the condition that anyone who consults it is

understood to recognise that its copyright rests with its author and that no quotation

from the thesis and no information derived from it may be published without the prior

written consent of the author.

This thesis may be made available for consultation within the University Library and

may be photocopied or lent to other libraries for the purposes of consultation.

Signature of Author . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Zhuangzhuang Dai

Acknowlegement

My sincere thanks to all the people who give me support during my pursuit for

PhD.

My deepest gratitudes go to my supervisors, Dr Robert Watson and Dr Peter Shep-

herd. Especially Dr Watson, I appreciate his scientific passion, profound knowledge and

experience which always guide me towards a right direction, and offering me an open

minded research environment. His dedication and commitment to research keeps in-

spiring me during my study.

I am grateful the Electrical and Electronic Engineering department of University

of Bath supports me financially, and the faculty and colleagues have been helpful and

patient.

I want to express my thankfulness to my parents who love and care about me

unconditionally.

I lovingly dedicate this thesis to my wife, Dr Xuan Wang.

Abstract

This thesis investigates a propagation modelling approach to navigation and source

localization. Complex urban propagation environments give rise to severe multipath

which impairs the reliability of conventional satellite and terrestrial based localization

systems. The motivation is the development of a location determination scheme ex-

ploiting multipath propagation. In this thesis a new ray tracing method has been

developed to efficiently determine channel characteristics. Using a database of channel

characteristics, a model is constructed to determine location of a receiver based on a

matching algorithm. This thesis also investigates into the inverse problem, i.e., the

source localization in urban environments. These methods have been tested against

noise and perturbations and is shown to be robust. In the simulated urban environ-

ments, navigation errors are typically less than 15m. The proposed source localization

algorithm is able to locate a radio source to better than 100m.

To evaluate these algorithms a 2.5D ray launching model has been developed, which

is able to make use of existing digital map databases. The accuracy of this model has

been validated by comparing simulated Received Signal Strengths (RSS) against chan-

nel sounding measurements in the city center of Munich, Germany. The ray launching

model makes use of parallel computing techniques using Graphic Processing Units

(GPU).

Key to the success of these methods is a new fingerprinting technique which corre-

lates abstract electromagnetic features with physical coordinates. This takes advantage

of data mining and machine learning to study patterns in the fingerprint distribution.

The thesis details the implementation of the navigation and source localization algo-

rithms. In particular, consideration is given to the problem of source localization using

a small Unmanned Aerial Vehicle (UAV). In order to efficiently solve this problem, the

technique of Dynamic Time Warping (DTW) has been explored. All errors have been

quantified and their sensitivities are determined.

ii

Contents

Acknowlegement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1 Introduction 1

1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Localization system decomposition . . . . . . . . . . . . . . . . . . . . . 5

2 Literature review 9

2.1 Traditional localization and signals of opportunity . . . . . . . . . . . . 9

2.2 Propagation modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 Digital mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.4 Fingerprinting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.5 Artificial neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.6 Pattern recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.7 Parallel computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 Ray launching model 26

3.1 3D modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3.2 Environmental entries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2.1 Antennas and propagation media . . . . . . . . . . . . . . . . . . 28

3.2.2 Urban layout entry . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2.3 Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.3 Propagation primitives . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.3.1 Reflection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.3.2 Diffraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

iii

3.3.3 Wall penetration . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3.3.4 Roof-top diffraction . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.4 Field reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.5 2.5D ray launching model validation . . . . . . . . . . . . . . . . . . . . 47

3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4 Navigation 53

4.1 Location fingerprint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.2 Localization algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.2.1 Mapping function . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.2.2 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.3 Accuracy and reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.3.1 Number of opportunistic sources . . . . . . . . . . . . . . . . . . 69

4.3.2 Location fingerprint components . . . . . . . . . . . . . . . . . . 69

4.3.3 Location target resolution . . . . . . . . . . . . . . . . . . . . . . 70

4.3.4 ANN configuration . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.3.5 Noise level during measurement . . . . . . . . . . . . . . . . . . . 73

4.3.6 Environmental factors . . . . . . . . . . . . . . . . . . . . . . . . 74

4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5 Source localization 78

5.1 Source localization algorithm . . . . . . . . . . . . . . . . . . . . . . . . 78

5.1.1 Data collector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

5.1.2 Feature transformation . . . . . . . . . . . . . . . . . . . . . . . 81

5.1.3 Source location fingerprint . . . . . . . . . . . . . . . . . . . . . . 82

5.1.4 Matching algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 83

5.2 Accuracy of analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

5.2.1 Measurement error and path deviation . . . . . . . . . . . . . . . 87

5.2.2 Noise level and bias . . . . . . . . . . . . . . . . . . . . . . . . . 92

5.2.3 Route selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

5.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

6 Enhancement and program optimization 107

6.1 Sequence-based navigation scheme . . . . . . . . . . . . . . . . . . . . . 108

6.2 Matlab . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

6.3 GPU-based acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

iv

7 Summary and future work 119

7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

7.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

v

List of Figures

1-1 Flowchart of navigation system . . . . . . . . . . . . . . . . . . . . . . . 6

1-2 Flowchart of source localization system . . . . . . . . . . . . . . . . . . . 7

2-1 Schematic diagram of satellite navigation . . . . . . . . . . . . . . . . . 10

2-2 Ray tube . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2-3 Bounding box division . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2-4 Google Earth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2-5 NVIDIA GTX 280M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3-1 3D ray launching model . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

3-2 Radiation pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3-3 Overpass Turbo panel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3-4 Propagation paths at the Arc de Triomphe . . . . . . . . . . . . . . . . 31

3-5 Transmission through a tree . . . . . . . . . . . . . . . . . . . . . . . . . 33

3-6 Reflection geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3-7 Diffraction geometry on edge . . . . . . . . . . . . . . . . . . . . . . . . 36

3-8 Transmission geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3-9 NYC street canyon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3-10 Roof top diffraction geometry . . . . . . . . . . . . . . . . . . . . . . . . 41

3-11 Reception detection mechanism . . . . . . . . . . . . . . . . . . . . . . . 43

3-12 Enclosed box scenario with seven objects inside . . . . . . . . . . . . . . 45

3-13 Power delay profile . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3-14 Angular Spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3-15 Measurement campaigns in Munich . . . . . . . . . . . . . . . . . . . . . 47

3-16 Pathloss measurements on Munich scenario . . . . . . . . . . . . . . . . 48

3-17 Simulated RSSs along ‘METRO 200’ . . . . . . . . . . . . . . . . . . . . 49



3-20 Scatter plot between simulated and measured pathloss along ‘METRO202’ 52

vi

4-1 Propagation paths in virtual scenario . . . . . . . . . . . . . . . . . . . . 55

4-2 PDP at NLOS receiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4-3 Angular Spread at NLOS receiver . . . . . . . . . . . . . . . . . . . . . . 56

4-4 Fingerprints generated by opportunistic sources . . . . . . . . . . . . . . 58

4-5 RSS heat map of a virtual scenario . . . . . . . . . . . . . . . . . . . . . 60

4-6 Artificial Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . 61

4-7 Hoxton district layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4-8 RSS heat map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

4-9 AOA x-axis Cartesian projection distribution . . . . . . . . . . . . . . . 64

4-10 AOA y-axis Cartesian projection distribution . . . . . . . . . . . . . . . 65

4-11 Neural Network user interface . . . . . . . . . . . . . . . . . . . . . . . . 66

4-12 Neural Network settings . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

4-13 Neural Network configuration . . . . . . . . . . . . . . . . . . . . . . . . 67

4-14 Training result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4-15 Mean error versus number of sources . . . . . . . . . . . . . . . . . . . . 69

4-16 Mean error versus target resolution . . . . . . . . . . . . . . . . . . . . . 71

4-17 Localization accuracy versus noise strengths . . . . . . . . . . . . . . . . 73

4-18 Localization accuracy versus measurement biases . . . . . . . . . . . . . 74

4-19 Tests A and B illustration . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4-20 Tests C bus insertion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

5-1 Static vs. dynamic remote sensing . . . . . . . . . . . . . . . . . . . . . 79

5-2 Dat collector: UAV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

5-3 Anticipated waveforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

5-4 Simulated UAV route . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5-5 Dynamic Time Warping versus Euclidean matching . . . . . . . . . . . . 85

5-6 Route deviation cases (a) and (b) on Hoxton scenario . . . . . . . . . . 88

5-7 Route deviation cases (c) and (d) on Hoxton scenario . . . . . . . . . . 89

5-8 Fingerprint characteristics of ‘METRO 200’ . . . . . . . . . . . . . . . . 91

5-9 Normalized accuracies versus noise strengths . . . . . . . . . . . . . . . 92

5-10 Normalized accuracies versus biases . . . . . . . . . . . . . . . . . . . . . 93

5-11 Mean error distance versus noise strengths . . . . . . . . . . . . . . . . . 93

5-12 Mean error distance versus biases . . . . . . . . . . . . . . . . . . . . . . 94

5-13 Accuracy versus biases after swapping coordinates . . . . . . . . . . . . 95

5-14 Mean error distance versus biases after swapping coordinates . . . . . . 96

5-15 Routes of measurement in Munich . . . . . . . . . . . . . . . . . . . . . 97

5-16 Pathloss Measurements on Munich scenario . . . . . . . . . . . . . . . . 98

vii

5-17 RSS heat map on Munich scenario . . . . . . . . . . . . . . . . . . . . . 99

5-18 Measurement routes against RSS heat map of Munich scenario . . . . . 100




5-22 Localization error versus noise strengths and bias . . . . . . . . . . . . . 103

5-23 Fingerprint of ‘METRO 200’ . . . . . . . . . . . . . . . . . . . . . . . . 104



6-1 Sequence-based navigation method . . . . . . . . . . . . . . . . . . . . . 108

6-2 A flowchart of master program . . . . . . . . . . . . . . . . . . . . . . . 110

6-3 Matlab profiling report . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

6-4 GPU memories and multiprocessors . . . . . . . . . . . . . . . . . . . . 113

6-5 Computation time versus CPU/GPU and scenario complexity . . . . . . 115

6-6 Visual Profiler results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

6-7 Visual Profiler zoomed-in on one iteration . . . . . . . . . . . . . . . . . 116

6-8 GPU usage analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

viii

List of Tables

3.1 Tree loss model parameters . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.2 Model validation against Munich measurement campaign . . . . . . . . 51

4.1 Emitter allocations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4.2 Transfer functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4.3 Mean errors against fingerprint primitives . . . . . . . . . . . . . . . . . 70

4.4 Accuracy versus number of neurons in each layer . . . . . . . . . . . . . 72

4.5 Accuracy versus training gain . . . . . . . . . . . . . . . . . . . . . . . . 72

4.6 Training specifications and performances on different scenarios . . . . . 75

4.7 Moving bus interference test . . . . . . . . . . . . . . . . . . . . . . . . . 77

5.1 Source localization case studies . . . . . . . . . . . . . . . . . . . . . . . 86

5.2 Accuracy performance versus deviations and measurement errors . . . . 87

5.3 Accuracy against Width of adjustment window . . . . . . . . . . . . . . 96

6.1 Computation time analysis . . . . . . . . . . . . . . . . . . . . . . . . . 114

ix

List of Abbreviations

ANN Artificial Neural Network

AOA Angle of Arrival

AWGN Additive White Gaussian Noise

BER Bit Error Ratext

CDF Cumulative Distribution Function

CPU Central Processing Unit

CUDA Compute Unified Device Architecture

DFT Discrete Fourier Transform

DTW Dynamic Time Warping

FDOA Frequency Difference of Arrival

FEM Finite Element Modelling

GNSS Global Navigation Satellite System

GPS Global Positioning System

GPU Graphics processing units

GTD Geometric Theory of Diffraction

IoT Internet of Things

LOS Line of sight

LPF Low-pass filter

MED Mean Excess Delay

MoM Method of Moments

MSE Mean Square Error

NLOS Non-line of sight

PDP Power Delay Profile

RF Radio Frequency

RFID Radio Frequency Identification Devices

RMSDS Root-mean-square delay spread

RSS Received Signal Strength

SHF Super High Frequencies

SNR Signal to Noise Ratio

TDOA Time Difference of Arrival

UAV Unmanned Aerial Vehicle

UTD Universal Theory of Diffraction

WGN White Gaussian Noise

Chapter 1

Introduction

Wireless communication has flourished in the past decades. With growing demands

on mobile telecommunication and internet connectivity, wireless communication links

are multiplying on the ground, in the air and on the sea. Since lower radio frequencies,

below 2GHz, are becoming crowded, capacious bandwidths at higher frequencies are ex-

pected to support downloading, video and audio streaming with improved data rates.

Empirical path loss models are rarely completely satisfactory in multipath environ-

ments, especially in urban areas where buildings and structures introduce complicated

propagation behaviour. On the other hand, this multipath interference directly results

from the propagation environment which can be exploited to provide geographical in-

formation through appropriate interpretation of the channel characteristics.

Traditional localization methods, e.g., satellite navigation systems, are mostly based

on a line-of-sight (LOS) propagation assumption. In this thesis, we are inspired to mit-

igate blind spots of traditional navigation by creating an urban positioning approach.

It is well recognized that deterministic methods for wave propagation modelling are

among the best methods for the estimation of radio channel characteristics. With the

aid of environmental specifications (e.g., building position and materials), determinis-

tic propagation models, such as ray tracing, imaging theory and Method of Moments

(MoM), can provide attractive channel response simulations that compare well to mea-

surement. Such models could guide off-line solutions to coverage prediction, radio

planning, navigation and source localization.

There is current interest in the determination of position and navigation in situa-

tions where satellite navigation signals from GNSS (Global Navigation Satellite System)

are denied. The denial of the GNSS signals may be via some form of intentional signal

jamming, unintentional through interference or merely a consequence of the challenging

environment, e.g., tall buildings in ‘urban canyons’. In these circumstances knowledge

1

of the propagation environment can help localize the source of the interference or alter-

natively be used in conjunction with signals of opportunity to aid a navigation solution.

Preliminary work of this thesis focuses on the development of a wave propagation

model suitable for urban environments. Then we aim at establishing mobile and source

localization algorithms through machine learning of the channel characteristics with

respect to location.

A 2.5D ray launching model was designed, implemented and tested by the author,

which is shown to be adequately efficient and accurate for the development of nav-

igation and source localization algorithms. It utilizes digital map databases, which

contain building and tree coordinates, in conjunction with semi-deterministic propa-

gating engines, including specular reflection, vertical edge diffraction, wall penetration

and roof-top diffraction. Power Delay Profile (PDP) and Angular Spread are the de-

liverables from the reception capture.

Feature eigenvectors, also termed ‘fingerprints’, can be extracted from the chan-

nel modelling results to assist training and learning of their geographical distribution

pattern. This leads to a mobile localization algorithm via Artificial Neural Network

(ANN) which are powerful in dealing with discontinuous high-dimensional data sets

such as these. The main objectives of this thesis are the generalization of such learning

processes on different urban scenarios as well as validation of its stability and reliability.

Source localization is an ill-posed inverse problem which involves a more careful

definition of fingerprint. We propose to use an Unmanned Aerial Vehicle (UAV) as a

probe to capture variations of trends of electromagnetic parameters along a specified

path. The outstanding frequency components are taken as fingerprints to check against

simulated source location candidates. A distance measuring algorithm, called Dynamic

Time Warping (DTW), is adapted to determine the best matched source location. This

method is shown to be accurate and robust.

This thesis aims to establish a reliable navigation and source localization system

based on propagation modelling in conjunction with fingerprinting and pattern recog-

nition techniques. The accomplishment of this objective requires exploration of the

questions below:

� What difficulties confront current navigation systems in urban environments?

� Is there any reliable and efficient radio source localization method in urban sce-

narios?

� What is the best approach to construct a navigation and source localization

system?

2

� What are the key requirements in order to exploit channel characteristics to aid

a positioning solution?

� What is the accuracy performance of a deterministic wave propagation model in

urban environments?

� What methods can be used to enhance the speed of deterministic wave propaga-

tion models?

� What are the accuracy and reliability of such navigation and source localization

methods?

In order to address these questions, this thesis is organized as follows (see also

Figures 1-1 and 1-2). The remainder of this chapter expands the underlying founda-

tion and motivation for this research. Chapter 2 reviews previous technologies and

researches existing urban localization methods, and explains their strengths and weak-

nesses. Various wave propagation models are investigated and a decision is made that

ray launching models are sufficiently reliable in generating channel characteristics. Un-

restricted access to digital maps is also demonstrated via the OpenStreetMap project.

Current techniques of data mining and machine learning are investigated which form

the basics of fingerprint extraction. This chapter also reviews multi-thread acceleration

methods for ray tracing.

Chapter 3 presents the development of our 2.5D ray launching model in detail,

containing antenna simulation, map abstraction, propagation mechanisms, reception

detection and validation. Chapter 4 explains the definition of location fingerprint and

how we utilize ANN to generate a mobile localization algorithm. The neural network

obtained exhibits convincing reliability and accuracy against noise and data corrup-

tions. Chapter 5 demonstrates the definition of source location fingerprint with the aid

of UAV measurements, and also the development of a matching algorithm. The stabil-

ity of the source localization method is tested against a series of perturbations, such

as route deviation, wind conditions, background noise strength and bias etc. where

satisfactory results are found. Acceleration and optimization methods are discussed in

Chapter 6. Finally, a summary and future work are presented in Chapter 7.

1.1 Background

Modern life is heavily reliant upon satellite navigation. From a military defence

system to daily driving, GNSS systems, such as the Global Positioning System (GPS),

are playing an important part. However, the propagation between satellite and ground

3

terminal can be vulnerable to multipath, bad weather or intentional GPS jamming.

Multipath propagation, especially in densely built areas, undermine the stability of

satellite navigation due to near ground reflections. Unpredictable weather conditions

and atmospheric properties bring about complex refraction and scattering into the

air-to-ground propagation. Alternative user localization plans are needed in such cir-

cumstances to make up for the weakness of GPS.

Jamming devices are radio frequency (RF) transmitters that intentionally block,

jam, or interfere with lawful communications, such as cell phone calls, text messages,

GPS systems and Wi-Fi networks [123]. Such jammers are illegal to use but unfortu-

nately very easy to buy and cheap to build. The consequence of GPS jamming can

be severe, e.g., allowing suspects to hide criminal activities. In this case, being able

to locate the jamming transmitter becomes a focus for security. Source localization

in multipath environments has been a long-standing problem, because triangulation

calculations cannot be directly used on corrupted non-line of sight channels. Previous

researchers have investigated whether multipath information can be exploited to aid

a positioning solution. Results testify the possibility as long as channel response is

clearly interpreted and a reliable localization algorithm is drawn.

The wireless communication technique has seen significant development in the past

decades, which brings great interest in comprehending the wave propagation charac-

teristics. As a result, channel response, radio location, as well as base-station planning

are all resolvable through effective propagation modelling instead of laborious measure-

ments.

Propagation modelling has been constantly evolving from statistical modelling of

propagation distance to deterministic 3D modelling for specific environments. This

technique is becoming mature in both efficiency and reliability. With detailed specifi-

cations of the propagation environment, there is growing reason to trust wave propa-

gation models to take the place of measurement campaigns which are both costly and

subject to errors due to temporal changes.

Data mining and machine learning, including the recent hot-spot ANN, enable

intelligent, adaptive and rapid generation of classification or regression functions. In

particular, pattern recognition could be applied to linking channel characteristics with

locations to formulate a localization algorithm. The off-line processing relies on a large

collection of training resources which can be extracted from the channel modelling.

This feature extraction based on physical characteristics, works in conjunction with

fingerprinting, a numerical transformation which quantifies abstract notions as vectors

or a matrix. A successful definition of fingerprint and an appropriate mapping function

would make a robust real-time localization approach.

4

The ever-growing computational capacity of digital processors, the wide-spread use

of mobile transceivers and easy access to digital city maps altogether create a solid

foundation for the formulation of the localization procedure. In terms of hardware,

cell phones, sensor networks as well as mobile carriers are able to fulfil cheap data

collection. State-of-the-art multi-core processors guarantee an efficient data analysis

power, even on embedded mobile devices. From a software perspective, a suitable

channel modelling tool, which is currently a relatively mature technique, and digital

map databases offer comprehensive urban simulation capability which can generate a

large amount of electromagnetic characteristics data.

These observations and the encouraging initial results of previous researchers, as

well as many industries inspire the investigation of this novel urban localization ap-

proach, which overcomes drawbacks of traditional navigation systems particularly in

multipath environments. From an engineering point of view, the required advanced

hardware facilities and software platforms are almost in place and the cost of such

systems continues to reduce.

1.2 Localization system decomposition

Through a comprehensive review of the literature in Chapter 2, we conclude that

deterministic wave propagation modelling continues to be a state-of-the-art channel

estimation technique, thanks to the increasing power of modern processors and digiti-

zation of map databases. The extensive use of wireless communication links provides

valuable information on channel characteristics in the context of Internet-of-Things

(IoT) and Big Data. We are, thus, exploring the possibility of digging and mining this

data from an information perspective. A ray launching model is developed to gener-

ate such channel characteristic data, which is demonstrated in Chapter 3, to take the

place of laborious measurement. Based on our research of previous measurements and

validations, we believe 2.5D ray launching models are sufficiently accurate and efficient

in urban environments. Acceleration and optimization of the ray launching program is

detailed in Chapter 6.

In this research we propose to investigate the feasibility of a model aided navigation

and source localisation approach. The basic hypothesis is that for a given environment

and a given source location, the trends in key signal characteristics such as Received

Signal Strength (RSS), Time of Arrival (TOA) and Angle of Arrival (AOA) as deter-

mined at various points throughout the environment, will be sufficient to determine its

location. This hypothesis happens to share a common idea with the ‘fingerprinting’

technique which identifies a target by extracting distinctive features and exhaustively

5

searching for the best match.

Therefore, we implemented a navigation system based on the extraction and recog-

nition of ‘location fingerprints’. As shown in Figure 1-1, the fingerprint database is

built from propagation modelling results with reference to the environmental entries

(building location and materials). Given the radio source and a grid of reference loca-

tions, the 2.5D ray launching model is used to generate ‘truth data’ for these locations.

Then, we adopt an ANN to help construct a deterministic mapping function between

the location fingerprints and geographical locations. Integration and evaluation of this

navigation process is expanded in detail in Chapter 4. A sequence-based navigation

scheme based on the fingerprint algorithm is introduced in Chapter 6 to improve the

off-line utility of the navigation system.

Figure 1-1: Flowchart of the navigation system with respect to content location.

If one considers a ‘Manhattan-style’ city layout, it is clearly possible to determine

6

the location of a source in an arbitrary location if one traverses all possible streets and

selects the largest RSS. However what is not clear is if one can derive generalised rules

for selecting the shortest trajectory of say a receiver payload aboard a small UAV to

achieve acceptable accuracy. This is a novel combination of both an inverse problem

and a propagation problem.

The core of a source localization system is to estimate, a-priori for a given known

environment, the minimum trajectory of receiver measurement locations in order to

best locate the source. This also encourages the discrimination of ‘source location

fingerprints’ as long as the UAV observations differ from each source location to another.

Figure 1-2: Flowchart of the source localization system with respect to content location.

As a result, we are inspired to define a ‘source location fingerprint’ based on UAV

data collection along a specified path in the street canyons. Procedures of this source

localization approach are shown in Figure 1-2. In this, the ray launching model plays

7

a significant role in generating UAV observations along the path with respect to source

reference locations. As the number of source reference locations increases, the potential

precision of localization increases, whereas, the price is a much higher computational

cost. From this data pattern recognition techniques are adopted to estimate the source

location as detailed in Chapter 5. The Discrete Fourier Transform (DFT) and Dynamic

Time Warping (DTW) algorithms are combined to implement the fingerprint matching

algorithm. The source localization performance as a function of various errors and

perturbations, as well as route selection is further investigated in Chapter 5.

In comparison to GPS (Global Positioning System) techniques the proposed mobile

and radio source localization approaches can employ lower frequencies (e.g., VHF/UHF

for DAB/DVB compared to L-band for GNSS), the use of which greatly reduces the

effects of building blockage while purposefully exploiting multipath propagation. A key

advantage over conventional fingerprinting methods is the ability to generate naviga-

tion information without the need to make detailed and costly measurements of the

environment.

To obtain fingerprints from modelling is cost effective and more flexible. If the

environment is changed due to, e.g., building construction or destruction, a set of fin-

gerprint database can be regenerated within hours using our acceleration techniques

as is illustrated in Chapter 6. Moreover, case studies and test results confirm that

our localization approaches possess a considerable degree of robustness in the pres-

ence of noise and interference. This suggests that an absolutely precise environmental

description is not necessary.

Note that in this work we assume a radio source is always static in terms of time

and position during measurement. Within the area of interest, it is regarded as the

only transmitter such that there is no cross-interference with other emitters. These

opportunistic signals are deemed to be narrow-band such that frequency dispersion

is not a concern. Last but not least, we assume that the simple devices pervasively

equipped on user mobiles and the UAV, are only able to measure values of RSS, TOA

and AOA with limited accuracy.

8

Chapter 2

Literature review

This chapter provides a comprehensive review of propagation and localization re-

lated literature. Conventional satellite based and terrestrial localization methods are

both found to be less effective in densely built areas mostly due to multipath propaga-

tion. In recent decades wireless communication links have seen remarkable growth in

distribution density as well as frequency range. This allows the possibility to use signals

of opportunities received by mobile devices which could be exploited for localization

use. To do this, it is crucial to establish a channel characteristics database in terms of

geographical location before an on-line localization method can be established. Catego-

rization and features of different propagation models are investigated. It is concluded

that a 2.5D ray launching model is a reliable and efficient tool for urban scenarios

which are accessible from rich digital map resources. Fingerprint-based data mining

and machine learning techniques are reviewed in pursuit of a reliable function bridging

the channel characteristics with physical locations. State-of-the-art parallel computing

methods using general purpose graphic cards have been explored to investigate the

capability of executing rapidly the propagation modelling stage.

2.1 Traditional localization and signals of opportunity

Currently, satellite navigation systems are the most widely used localization ap-

proach [49]. A built-in electronic receiver for the satellite constellation system is neces-

sary to achieve autonomous geo-spatial positioning. The local time of the receiver can

also be determined with high precision to accomplish synchronization. Global Navi-

gation Satellite System (GNSS) refers to any satellite navigation system with global

coverage. Up to now, mature GNSSs include the European Union’s Galileo system,

the Russian GLONASS, China’s BeiDou system and the United States NAVSTAR

9

Global Positioning System (GPS). The GPS system provides the broadest application

for military, civil and commercial use. The GPS system does not require the user to

transmit any data, and operates independently of any telephonic or internet reception

[31], although the usefulness of the GPS positioning can be improved by these assistant

functions. A schematic diagram of satellite navigation application is shown in Figure

2-1.

The accuracy of GNSS localization is typically very high, usually less than 5m [16].

Nevertheless, line-of-sight (LOS) propagation between satellite and receiver is critical

because theoretically time delays from at least three satellite signals are required to

allow positioning to an exact point. However, this hypothesis does not always hold in

densely built urban areas [125]. Flat building surfaces typically act as reflectors cre-

ating receiver images which undermine the satellite localization precision [29]. Severe

multipath propagation in urban environments also makes synchronization problem-

atic. Moreover, the satellite-to-ground propagation path may be distorted merely as a

consequence of bad weather.

Figure 2-1: Schematic diagram of satellite navigation on ground, in ocean and air,taken from: https://www.laboratoryequipment.com/news/2013/03/ground-system-improves-satellite-navigation.

Terrestrial communication links can also be used for geo-positioning. Such methods

rely on a trilateral geometric estimation which requires three mobile link measurements

10

of TOAs or AOAs. Such terrestrial localization is also affected by multipath from high-

rise typical of modern cities. Furthermore, TOA and AOA observations may not be

due to LOS propagation as expected.

Most traditional localization methods lack reliability in urban localization scenarios

which has seen growing interest in recent decades. Many attempts have been made

to exploit knowledge of the propagation environment in conjunction with signals of

opportunities to aid the localization solution. Schmidt [108] made exploratory research

into the diversity of signal parameters which can be used to locate radio sources. The

usability of RSS, TOA and AOA in terms of indoor localization are discussed in [14,

19, 24] where good accuracies are found through experiments. In [63, 83], the authors

developed an efficient ray tracing tool for indoor localization. However, these models

are too simplistic to see wide application for complex urban scenarios.

Huang et al. [45] proposed the use of weighted averaged RSSs of wireless commu-

nication links for indoor navigation. This approach presented very good reliability in

the presence of measurement noise. Schmitz et al. [110] tested the feasibility of using

TDOA (Time Difference of Arrival) as an indicator to localize mobile users in urban

environments. The merit of this approach is that it purposefully takes advantage of

the multipath propagation. However, reliability tests when subject to noise and inter-

ferences are not given. Del Corte et al. [22] put forward a source localization approach

which only requires several AOA measurements without any knowledge of the environ-

ment. A key requirement is the determination of the dominant ray and its delay, which

is problematic. Kelner et al. [55] attempted to locate the emitter based on the signal

Doppler frequency. However, little validation for different environments is given.

All of the previously described localization methods have been developed using

signals of opportunity. A signal of opportunities can be considered to be a signal

transmitted for non-navigation purposes but from which time of flight and other useful

information can be determined [52, 87, 103]. In urban environments, the growing

number of wireless communication links, e.g., Wi-Fi sources, offer rich resources of

opportunistic signals with spatial diversity. Collecting and comprehending these signals

could not only help localize mobile users but reveal unknown source locations through

an inverse approach.

The study of the complete propagation channel including antenna radiation pat-

tern is the key to successfully comprehending opportunistic signals. Knowledge of the

antenna pattern is well known for the types of antenna typically used in mobile user

equipment. Taking mobile phone for example, the antennas range from a planar in-

verted F antennas, a planar meander line, a folded loop to a modified dipole. The

major part of channel reconstruction relies on multipath propagation modelling which

11

will be discussed in the next section.

2.2 Propagation modelling

The physical layer properties of communication systems have drawn increasing sci-

entific and industrial interest in the past few decades. The lower frequency bands of the

microwave range have been well studied and pervasively used [102]. A major advantage

of using lower frequency signals is that free space propagation loss and building pen-

etration losses are reduced [84]. Thus, it is feasible to broadcast over large distances

with limited transmitted power.

Traditional radio planning methods typically rely heavily on expensive as well as

time consuming field measurements. From these measurements, empirical wave propa-

gation models, such as Okumura-Hata, COST-231, Walfish-Bertoni models have been

developed to provide rapid path loss estimation in typical urban environments [12, 101].

However, many of these models fail when applied to other frequencies and scenarios.

Furthermore, measurement-based calibration may ultimately be required.

Since plenty of applications are saturating the lower RF bands, modern wireless

communications are seeking wider bandwidths, so as to increase data rate by migrat-

ing to the higher microwave frequency band [25]. For example, 2.45GHz ISM bands, 6

GHz and up to 60GHz bands are being investigated for 5G applications [42, 43, 102].

However, these signals suffer higher propagation loss and encounter more complicated

scattering phenomena [139]. Deterministic models are thus being developed. Full-wave

models, e.g., Method of Moment (MoM), can solve the wave propagation behaviour

based on Maxwell’s equations [48]. Although they provide good accuracy, the compu-

tational cost rises exponentially as the size of the environment increases.

Ray tracing and ray launching models based on geometrical approximations can also

be categorized as deterministic models [97]. They are state-of-the-art techniques for

indoor modelling thanks to the comprehensive inclusion of environmental details which

essentially account for multipath propagation. Both ray tracing and ray launching

methods use the concept of ray optics but employ different approaches.

Ray tracing is more of an objective-oriented simulation. It constructs images of

virtual sources or receivers with respect to object surfaces to identify potential path.

A visibility tree is often adopted to establish the network of consecutive reflections and

diffractions [7, 135]. A representative application of ray tracing is the corridor model

in which different orders of reflections bouncing along a tunnel are assumed to be the

potential propagation paths [15]. Corridor models are very effective and efficient in

indoor modelling and may be applied to urban street canyons [6].

12

The performance of ray tracing can be outstanding as long as detailed environment

information is available. However, the computational complexity is regarded as a weak-

ness of ray tracing even though multiple acceleration means have been proposed and

experimented, such as bounding boxes to mitigate redundant imaging [44], preprocess-

ing and discretization of the visibility tree [30, 74], leverage of multi-core processors

[107] etc. Acceleration performances show a strong dependency on the objective sce-

narios.

By comparison, the ray launching is a transmitter-oriented brute force simulation.

A large number of equally spaced rays are shot from the transmitter which represents

the wavefront [66, 131]. A reception detection algorithm is designed to capture the rays

being received [106]. Superposition of the arriving rays depicts the nature of construc-

tive or destructive interference. In the process, reflection, diffraction and transmission

behaviours of each ray is traced and recorded [64].

A ray is terminated if one of the fading-out criteria are satisfied. Ray launching

usually does not consider scattering because specular reflections on a surface form

another wavefront [10]. As long as the number of rays is adequate, different propagation

behaviours of a wavefront upon a structure can be determined.

A significant advantage of ray launching lies in the propagation paths generated by

an emitter being independent of the receiver. Therefore, an identical receiving mecha-

nism can be assigned across a scenario to obtain channel parameters with respect to all

receiver locations in one go. Furthermore, the intersection behaviours of rays during

propagation are also independent which allows acceleration using parallel computing.

In summary, although a brute force approach, better speed performance is anticipated

using ray launching.

According to [87], convergence experiments show that with over seven orders of re-

flections the ray launching estimations broadly agree with measurements. Alwajeeh et

al. [9] found that ‘10R1D’, meaning 10 orders of reflections and one order of diffraction,

were sufficient to deal with complex urban environments. As is known, the computa-

tional cost of ray tracing rises exponentially when the reflection order increases. Ray

launching maintains a constant number of wavefronts thus suffering less from the in-

crease of the reflection orders.

13

Figure 2-2: An illustration of ray tube intersection detection taken from [135].

The selection of model is indeed a trade off between expected accuracy and com-

putational complexity. So far, deterministic propagation models are more reliable and

accurate, and represent the future trend. Although a ray launching model is deemed

superior in most aspects, it requires further simplification to suit large area urban en-

vironments. Researchers in the literature have attempted to improve ray launching

accuracy and efficiency by applying tube-based ray shooting [100, 135]. This method

regards a group of neighbouring rays as a ray tube which should see similar behaviour

if intersecting with a common flat surface. Once a significant intersection geometrical

difference is found within a tube, the boundary of a surface is detected. Therefore,

the reflected wavefront can be defined by only considering the boundary tubes which

reduces intermediate computations. However, the computation of intersection criteria

is more complicated and even more dense rays are required. Consequently, the accel-

eration provided by ray tubes is questionable. An illustration of a ray tube is shown

in Figure 2-2.

Azpilicueta et al. [12] proposed the use of an Artificial Neural Network to recognize

the pattern of similar intersection behaviours, so as to speed up ray launching process in

indoor applications. Agelet et al. [7] proposed to divide a scenario into bounding boxes

to eliminate redundancy intersection searches as is illustrated in Figure 2-3. Although

good performance is found using these methods, they greatly depend on the scenario

being investigated.

14

Figure 2-3: Dividing a scenario into bounding boxes to reduce computation complexityon intersection detections taken from [7].

Five propagation primitives are usually specified in a full-wave deterministic model,

including free space propagation, reflection, diffraction, wall penetration and diffuse

scattering [11]. Excluding diffuse scattering which is not necessary in ray launch-

ing, some propagation behaviour can be extremely expensive to simulate but making

very limited contribution to the accuracy performance [68]. For example, modelling

wall penetration not only demands precise information on thickness and material, but

involves complicated polarization [86]. As a result, integrating semi-deterministic al-

gorithms into the deterministic model provides a reasonable simplification [136].

The Friis transmission equation is regarded as the most popular formula to estimate

the path loss as a function of propagation distance [105]. The received power, Pr,

calculated using this formula can be written as;

Pr =GtGrλ

2

(4πd)2· Pt (2.1)

where Gt, Gr are transmitting and receiving antennas gains, respectively, λ denotes the

wavelength, d is the cumulative travelled distance and Pt is the transmitted power. Al-

15

though phases of signals are omitted which results in a missing presentation of construc-

tive or destructive interference, the Friis transmission equation is adequately accurate

and extremely efficient in far field applications [102], for example urban propagation

modelling where the scale is much larger than multiple wavelengths of opportunistic

signals.

The reflection coefficient is defined to compute how much electromagnetic energy

is dispersed in that ray with respect to incidence angle and surface material [76]. In

terms of polarization, the vertically polarized component of the reflected ray is derived

from the vertical form of reflection coefficient;

Rvertical =ε · sin θ −

√ε− cos2 θ

ε · sin θ +√ε− cos2 θ

(2.2)

and the horizontally polarized component follows;

Rparallel =sin θ −

√ε− cos2 θ

sin θ +√ε− cos2 θ

(2.3)

where θ is the incidence angle, ε represents the electric permittivity of the surface

material which can be obtained from;

ε = εr − j60σλ (2.4)

where εr is the relative electric permittivity, σ indicates the electrical conductivity, λ

equals the wavelength.

As is widely acknowledged that UTD [33, 95], which stems from GTD [17, 54],

provides a precise description of diffraction. Diffraction can be categorized into edge

diffraction which generates a plane circle or cone-like diffracted scatterers, and pin point

diffraction in which the pin point becomes a virtual source and diffuses the energy all

over the space [60]. As a result, genuine ray launching modelling of diffraction means

thousands of hybrid rays would have to be considered. Owing to a high diffraction loss,

typically only one order of diffraction in urban environments is considered. Similarly,

researchers in the literature also defined ‘diffractivity’ [11] to calculate edge-diffracted

rays in a deterministic way.

In the case of long distance propagation, the most significant arrival is likely to

come from multiple roof-top diffraction [137], however, the exact roof-top diffraction

can be difficult to reconstruct. In order to ease the computation, a semi-deterministic

roof-top diffraction model is adapted in Chapter 3 to simplify our 2.5D ray launching

engine with little compromise in accuracy. These definitions of propagation mechanisms

constitute the core of a ray launching engine.

16

Common 2D propagation models employ digital map databases containing geo-

metrical data on objects, surface or edge coordinates and tree locations. Given the

transmitter position and power, a cluster of rays are initiated with a specified increment

angle. As each ray propagates in the environment its first incidence with surfaces, edges

or the boundary of the defined area will be processed with the help of either Fresnel

reflection equations [11], UTD or transmission models [69, 90]. Intersection detection

which determines incidence points will not stop iterating until the ray strength fades

below a threshold level, the intersection order reaches a specified limit, or shooting out

of the zone of interest.

A ray-capturing mechanism is usually adopted to assist arriving detection, e.g., a

reception sphere [106]. The electromagnetic field strength at the receiver is the sum

of each ray that passes through the receiver position, i.e., superposition of all arriving

rays.

3D wave propagation models could be extended from 2D models with selective

considerations of interactions as well as amendment of coordinate system [62, 70, 71].

Firstly, ray vectors should be defined at the transmitting point in angular representa-

tion with designated rotation angle and elevation angle increment. Then data about

object parameters should be stored, such as object positions, surfaces, edges or vertices

functions, material conductivity, permeability, dielectric constants etc. Via ray vector

intersecting with surfaces, edges or vertices functions, the travelled distance, incidence

position and angle to the object as well as signal strength can be derived. The processor

genre depends on which kind of incidence takes place, specular reflection, transmission,

edge diffraction, or vertex diffraction [28, 74].

3D models are widely used in indoor environments at high frequency bands be-

cause more sophisticated multipath effect is anticipated in the channel. However, the

computational cost is even higher as a price. An experimental 3D modelling attempt

in Chapter 3 verifies this statement. 2.5D models, which only consider vertical do-

main diffractions and ground reflections on a basis of 2D models, are generally deemed

as a practical solution for urban propagation modelling [65]. Literature reviews have

demonstrated effectiveness of 2.5D models in different urban scenarios [39, 41].

Overall, wave propagation modelling is able to present dynamic channel characteris-

tics as a function of spacial geometry. To be able to address the multipath phenomenon

plays a growingly important role in modern wireless communication applications. The

ever-increasing computation capability of processors and easy access to city layout

databases are encouraging the widespread use of deterministic models.

17

2.3 Digital mapping

A successful application of propagation modelling in cities relies on an accurate

environmental description. Nowadays digital map databases are prevailing for military,

commercial, academic as well as public uses. 2D digital maps are easily accessible.

Many government based and commercial institutions have developed comprehensive

3D building databases covering a vast majority of cities over the world. For instance,

Google Earth stores terrain data from satellite imaging which is able to present coverage

type, height above sea level as well as detailed shape of a building or structure. An

aerial view of Hoxton district in London is captured from Google Earth shown in Figure

2-4.

Figure 2-4: An aerial view of Hoxton district in London captured from Google Earth.

Coast initiated a digital map project called OpenStreetMap in 2004 [2]. It con-

structs a free editable map of the world from volunteered geographic information. With

over 2 million registered users, 2.5D data of most urban layouts, i.e., building latitude

and longitudes as well as heights, are available and constantly updated. Quite a few

researchers in the literature have adopted and validated the OpenStreetMap in the

use of urban propagation modelling [88]. An advantage of using the OpenStreetMap

for propagation modelling is that the database is constantly updating to approach a

real-time reconstruction. Moreover, it is free.

A powerful user interface, named as Overpass turbo [3], is built upon the Open-

StreetMap. It relies on a set of API queries to carry out data mining upon the in-

teractive map. Depending on the user requirements, Overpass turbo returns specific

18

geographical information, such as public, commercial or residential buildings, vegeta-

tions, parks, rivers, motorways, railways etc., in digital formats. For example, a kml

file can be generated and translated into structural arrays for Matlab. The object

extraction process is detailed in Chapter 3.

Rich and detailed digital maps with high accuracy give a solid foundation for prop-

agation modelling in urban environments. Compared to current indoor deterministic

modelling where environmental information has to be specially entered and yet con-

stantly changing, urban modelling remains effective for a longer period such that it

possesses better temporal validity for practical use.

2.4 Fingerprinting

Laborious data collection and dynamics of the urban environments can be effec-

tively addressed using deterministic propagation modelling. Given objects layout in

the scenario and electromagnetic coefficients of materials, a combination of RSS, AOA

and TDOA at known positions may provide evidence to predict locations in context of

multipath [40]. In fact, RSS has been widely investigated and validated in indoor local-

ization systems [18, 113, 132]. The major concern is that the number of opportunistic

signals needs to be adequate to support a robust accuracy of analysis. Measurements of

AOAs with coarse precision, e.g., using ’MUSIC’ method [23, 116], are also capable of

producing spatial information as long as multipath propagation is explicitly identified.

Jin et al. [51, 110] proposed and experimented on using TOAs to discriminate emitter

locations. Similarly, identification of the propagation path is the key to success.

A new concept called ‘fingerprinting’ in terms of channel characteristics has drawn

increasing attention recently. Uniqueness is the most significant feature of a fingerprint.

Also a fingerprint should possess a degree of tolerance such that slight distortions can

be mitigated. This requires appropriate quantification of the fingerprints. Since spatial

diversity has been broadly recognized in channel characteristics [34, 119], fingerprint-

ing may be an approach to finding an access point or reference point locations. In the

wider use of the term, a radio fingerprint can be defined as an eigenvector incorpo-

rating components such as RSS, AOA, TOA etc. Taking advantage of opportunistic

signals, the eigenvector may be expanded to incorporate diverse channel characteristics.

Given a reliable mapping function, such fingerprints might lead to correct geographical

locations.

Fingerprint-based localization proposals are presented in [8, 22], where dominant

AOAs, RSSs and TDOAs are registered as fingerprints. Very good accuracy was found

through tests. Zou et al. [141] developed a novel approach which extracts location

19

characteristics by combining wavelet transformation and singular value decomposition.

The eigenvectors derived are then used to build the fingerprint-location database. Sim-

ulation results exhibit very good localization accuracy.

Alternatively, unknown emitter locations may be revealed via fingerprinting because

an emitter is generating different field observations at different locations. Coco et al.

[20] proposed to find the most probable source location among potential grid points

by solving a cost function established upon measured fingerprints. Ray tracing is the

ideal tool for generating source fingerprints and good accuracy is seen from simulation.

Phelan et al. [92] extracted characteristics, such as SNR (Signal to Noise Ratio),

BER (Bit Error Rate), AOA and TOA, from channel to locate a radio source using a

clustering algorithm. This method is validated through experiments. However, the user

is required to obtain detailed information on what is being transmitted. Wadhwa et

al. [126] designed and tested an interesting tracking algorithm utilizing an Unmanned

Aerial Vehicle (UAV) to automatically approach the emitter in which RSS essentially

makes the fingerprint. Although the algorithm is validated and optimized, it may take

a long while for the UAV to arrive at the source.

2.5 Artificial neural network

A proper fingerprint definition is a priori to bond channel characteristics and loca-

tions, nevertheless, a robust matching algorithm referees the success of fingerprint-based

localization systems. Linearity certainly does not apply to high-dimension fingerprint

eigenvectors. Although non-linear regression methods are tested and proved working,

very accurate measurements on TOA or AOA are critical to perform a valid location

estimation [77, 103]. In this case, subtle uncertainties of temporal synchronization,

variations in antenna radiation pattern, constructive or destructive interferences bring

serious problems to the problem of localization accuracy.

In order to map high-dimensional discontinuous fingerprints with a mass of urban

positions, machine learning, particularly Artificial Neural Network (ANN), accounts

for a potential solution. Some researchers in the literature have succeeded in training

passive radio measurements using ANN [127, 141]. A basic hypothesis for this approach

is that the fingerprints extracted are unique and static to define locations.

ANNs are famous for handling discontinuous data. The more training resources

there are the more trustworthy a mapping function would be. ANNs are comprised of

neurons, connections and weights, propagation function and learning rule [1]. These

parameters are all liable to user customization for an optimized performance.

20

An ANN can be generally regarded as a simple mathematical map relationship;

f : X → Y

whereX stands for a set of training inputs, and Y is usually called target. Once function

f is agreed on convergence, it automatically generates recognitions on unknown input

materials. Depending on learning paradigm used, ANN is divided into three types,

including supervised learning, unsupervised learning and reinforced learning. Since

the fingerprints are essentially labelled channel characteristics, it is supervised learning

that best fits our application.

The core power of ANN comes from its ability to learn. The goal is basically min-

imizing a cost function which is chosen by the user. Back-propagation ANN is the

most widely acknowledged learning pattern when used in conjunction with fingerprint-

ing [124]. Back-propagation algorithm is a method to calculate the gradient of the loss

function (produces the cost associated with a given state) with respect to the weights in

an ANN. Sotiroudis [115] succeeded in mapping RSSs and TDOAs onto locations, using

feed-forward back-propagation ANN which marked a milestone for fingerprint-based lo-

calization. It is also noted that the ANN produced smooth outputs of estimation given

erroneous or out-of-bounds inputs which is ideal for localization purpose.

2.6 Pattern recognition

The reliability of ANN depends on a large pool of training resources. In terms of

radio source fingerprinting, it takes a great deal of simulation effort for a ray launching

model to obtain sufficient source fingerprints for training. In this case, applying pattern

recognition techniques on easily accessible fingerprints could make a more practical

solution.

Pattern recognition is a branch of machine learning that focuses on the recognition

of patterns and regularities in data [73]. Labelled training data, or so called supervised

learning, is usually the resource for pattern recognition. The processing on both labelled

and unlabelled data is referred to as data mining, which is the transformation process

of large data sets for better understanding involving machine learning, statistics, and

database systems. Data mining, particularly pattern recognition, using fingerprints

may be able to locate unknown RF sources. In [91], it can be seen that a diverse range

of engineering problems are currently seeking for answers from pattern recognition.

The K-Nearest Neighbour (KNN) is one of the simplest but most popular data min-

ing algorithms. Literally, k closest candidates are regarded as a cluster for classification

and regression purpose. KNN is a type of instance-based learning where the function is

21

only approximated locally and all computations are deferred until classification [133].

When k equals 1, an intuitively closest match in the feature space is returned. This

can be useful for source localization if the fingerprints are reasonably dispersed in the

feature space.

KNN algorithm can be adapted to a weighted classifier which assigns a measure of

distance to each of the k neighbours. Such classifier is also called K-Weighted-Nearest

Neighbour (KWNN). KWNN allows simplistic study on distribution pattern of dense

training samples. Huang et al. [45] used both KNN and KWNN as matching algo-

rithms for RSS-based indoor localization and verified significant improvements using

the latter over the former. Zou et al. [141] applied KWNN for recognition of extracted

electromagnetic fingerprints where very good accuracy was found.

Dynamic Time Warping (DTW) is an alternative pattern recognition algorithm

specially designed for time series to mitigate mismatching. Given two data series −→a =

(a1, a2, ..., an) and−→b = (b1, b2, ..., bn) of length n and m, respectively, an n-by-m matrix

can be constructed in which the (i, j) element contains the distance between two points

ai and bj . Each matrix element corresponds to the alignment between points ai and

bj . A warping path W is a contiguous set of matrix elements that define a mapping

between −→a and−→b . The k -th element of W is defined as wk = (i, j)k. Hence, the DTW

is the path that minimizes the warping cost given by

DTW(−→a ,−→b ) = min

{∑k=1

wk

}

If a complete spatial distribution of electromagnetic characteristics can be obtained,

radio source localization becomes a deterministic and simple problem. Since LOS short-

distance propagation always produces higher path losses than NLOS long-distance

propagation, the peak of RSSs distribution, the valley of TOAs distribution, or the

radiation center of the most significant AOAs distribution on an aerial-view 2D heat

map definitely represents the emitter location. As a result, it is straightforward to

trace the radio source. However, precise and dynamic real-time channel monitoring

of an urban scenario is not practical at all. Source localization essentially makes a

compressed sensing problem.

Compressed sensing aims at an efficient reconstruction on undetermined linear sys-

tems. This is based on the theory that, via optimization, the sparsity of a signal can

be used to recover it with far fewer samples than required by the Shannon-Nyquist

sampling theorem [122]. It requires merely a fraction of outstanding samples to reveal

the source location instead of the whole picture. This involves careful selection of the

sampling strategy. Walter et al. [128] utilized compressed sensing to locate unknown

22

emitters with good accuracy found. Whereas, the power supply and maintenance bring

new technical difficulties. Moreover, questions still remain where to put the sensors and

how many of them are necessary. These also motivate us to use a UAV as remote sensor

which is further explained in Chapter 5.

The various electromagnetic characteristics do not offer equally significant informa-

tion in terms of localization capability. These labelled features can be transformed onto

other labelled domains to reveal information from different perspectives. The Discrete

Fourier Transform (DFT) is a lossless transformation from time domain to frequency

domain. The k -th DFT coefficients of x, Xk, can be expressed as

Xk =N−1∑n=0

xn · e−j2πkn/N (2.5)

where N is the length of samples and xn represents the n-th sample value. Real

and complex parts of Xk exhibit periodic features of x.

Labelled features can also be transformed onto unlabelled domains to present ex-

plicit mathematical significances, e.g., using Principal Component Analysis (PCA).

PCA is a statistical procedure which uses an orthogonal transformation to convert a

set of observations of potentially correlated variables into a set of values of linearly

uncorrelated variables [117]. PCA is often used for he purpose of dimension reduc-

tion. Applying PCA on fingerprints may allow us to find the best coordinate system

to differentiate them although physical meanings would be lost.

There are many other pattern recognition instances for solving localization prob-

lems. Jin et al. [51] successfully adopted extreme learning machine upon RSS variation

gradient to predict radio source location. Coluccia et al. [21] integrated Maximum Like-

lihood into localization algorithms by minimizing a cost function which is transformed

from RSS-location correlation matrix.

2.7 Parallel computing

Parallel computing is currently a hot topic of study which breaks a large serial

task into sub-tasks and executes all concurrently. Parallelism can take different forms

such as bit-level, instruction-level, data-level, and task-level parallelisms. Multi-core

processors are the most commonly seen parallel computing application. For example,

the dual-core and quad-core CPUs (Central Processing Unit), which are widely used

in modern computers, allow much faster computations compared to a single core of

identical specifications.

Recently the computation potential of general purpose Graphic Processing Units

23

(GPU) is being recognized. The GPU is now far more than an embedded device for

display operations. Modern GPUs are equipped for double-precision mathematical op-

erations in a parallel configuration which sees broader applications in smart computa-

tion. Open Computing Language (OpenCL) and Compute Unified Device Architecture

(CUDA) are classic parallel programming platforms designed for GPUs. An NVIDIA

GTX 280M GPU is shown in Figure 2-5. OpenCL is a high level programming language

with comprehensive supporting libraries. However, the CUDA platform provides direct

access to the GPU virtual instruction set and parallel computational elements, which

is more convenient for senior users to customize kernel assignment [104].

Figure 2-5: NVIDIA GTX 280M Graphic Processing Units taken fromhttp://www.nvidia.co.uk/page/home.html

Although parallel computing is able to bring dramatic acceleration, it requires care-

ful time and resource management on both host and device. OpenCL and CUDA pro-

grams are more difficult to write than sequential ones because concurrency introduces

several new classes of potential software bugs [104], of which race conditions are the

most common. Communication and synchronization between the different subtasks

need significant work to get the best out of a parallel computing architecture.

As can be easily inferred, parallel computable subtasks should possess little depen-

dency on each other during processing such that threads do not always have to hold

for synchronization. There are actually multiple layers of independent computations

in propagation modelling. For example, the channel characteristics between transmit-

ter and receiver of different locations are derived individually, so are the propagation

paths of the launched rays. The matching between the fingerprint candidates and each

fingerprint in the database is also independent. Many researchers in the literature have

24

successfully applied GPU-based parallel computing to speed up wave propagation mod-

elling [82, 107]. Acceleration of performance using CUDA with our model is further

considered and demonstrated in Chapter 6.

2.8 Summary

This chapter provides a review of the existing propagation models and localization

methods, as well as their limitations. In order to overcome these limitations, state-

of-the-art propagation modelling techniques are then discussed on the possibility of

aiding a fingerprint based localization solution. Data mining and machine learning can

potentially be exploited to generate location decision algorithms. It is also noted that

parallel computing architectures might be used to accelerate such a new localization

scheme.

25

Chapter 3

Ray launching model

For the work in this thesis, a ray launching model has been developed and integrated

from scratch. It is based on Matlab 2016b and consists of a few thousand lines of

code. The model is expected to be able to simulate electromagnetic characteristics at

different locations with respect to a radio source in a specified propagation environment.

Although initially we considered a 3D ray launching model, our initial trial took a

tremendously long time to process. Thus, the focus of this work has become a 2.5D

model which proved to be robust in typical urban environments.

A traditional ray-type wave propagation model consists of three phases: initiation,

ray shooting and field reconstruction; although different ray-type models may have

more pre-processing steps [109]. The initiation phase aims to provide details on the

electromagnetic properties of the environment and set up the propagation geometry

of the transceivers. During the ray shooting phase, free space propagation applies to

each ray unless intersection occurs. Depending on which kind of intersection is taking

place, a reflection, diffraction or transmission processor is exploited to calculate the

loss and redefine the ray that propagates forwards. Concurrently, a reception detection

processor keeps a record of the rays that are arriving at the receiver. Eventually, the

received signal properties are acquired from the superposition of all rays that fall into

the reception sphere.

This 2.5D ray launching model has also been validated against COST 231 action

measurements in the city center area of Munich from Mannesmann Mobilfunk GmbH

[36]. The mean absolute error along any of the three measurement routes is less than

8dB, which proves the effectiveness of this model.

26

3.1 3D modelling

Urban propagation environments are usually very complex. Various factors lead

to a remarkable degree of uncertainty to accurately predicting the ray propagation

behaviour, e.g., irregular geometric shapes of modern architectures, diverse surface

materials, different kinds of vegetation, traffic and pedestrians [102]. A useful prop-

agation modelling tool should be able to address the representations of these factors.

Therefore, the first step is to decide the genre of modelling.

Figure 3-1: An attempt of 3D ray launching modelling in which buildings and groundare divided into triangles and coloured in yellow; blue lines represent LOS propagationpaths or reflections; red lines represent diffracted rays; green dots are intersectionpoints; magenta arrows represent rays terminated or without reception.

As has been discussed in Chapter 2, 3D ray launching models are among the most

powerful deterministic models. We attempted to build a 3D ray launching model

based on Matlab. A demo of the propagation paths in a virtual environment is shown

in Figure 3-1. The ground is represented by a square plane of 500m side length. We

defined four cubic buildings and a triangular pyramid with flat surfaces. A radio source

is allocated at (200m, 200m, 20m) transmitting evenly distributed rays over a spheric

surface. Note that all flat surfaces are split into triangles because they are easy for

ray-surface intersection computations.

27

It can be seen that the ray coverage becomes sparse as they travel away from the

source. This would likely result in missing intersections. The only solution is to increase

the density of rays being shot which enormously increases the computational cost. Even

though thousands of rays are launched from the source in this demo taking over 20

minutes to run, the 3D model detects few edge diffractions with close-by buildings.

We suppose that hundreds of thousands of rays are necessary for the comprehensive

modelling of urban scenarios. Such large computation time for only a modest increase

in accuracy means that a 3D approach is not useful in this study. Hence, we determined

to build a 2.5D model. The development process is detailed in the rest of this chapter.

3.2 Environmental entries

A detailed description of the environment and measurement devices is critical to

the success of a wave propagation model. As has been stated previously, comprehensive

digital maps of most cities are easily accessible. The urban propagation environment

should be as fully defined as possible to ensure a close-to-reality propagation prediction.

This section explains the preliminary works in the initiation stage.

3.2.1 Antennas and propagation media

Before initiating the ray launching model, the antennas and propagation media

have to be specified. The transmitter and receiver locations also need to be known as

a priori in order to perform reliable channel estimation. The antennas patterns can be

generated and encapsulated as a function of departing or arriving angle. An example

of a 2D radiation pattern is shown in Figure 3-2.

The media in urban scenarios can generally be regarded as air or free space. And

the propagation in these environments usually satisfy far-field propagation condition

[35]. This type of propagation within a homogeneous media always follows a straight

line. Therefore, the Friis Transmission Equation can be applied to approximate the

path loss due to free space propagation [105]. The derivation of received power, Pr, is

given by Eqn. (3.1)

Pr =GtGrλ

2

(4πd)2· Pt (3.1)

where Gt, Gr are the transmitting and receiving antennas gains, respectively, λ is the

wavelength, d is for the cumulative travelled distance and Pt is the transmitted power.

Hence, the total travelled distance, departing angle from transmitting antennas and

arriving angle at the receiving antennas should be recorded for each ray in order to

28

give a general path loss estimation. Since the wavefront is represented by a finite

number of rays, the transmitted power of each ray is obtained from splitting the source

power into the number of rays being shot.

Figure 3-2: An example of antennas radiation pattern in horizontal domain, the dashedred envelope represents the gain as a function of angle in horizontal plane.

3.2.2 Urban layout entry

As has been introduced in section 2.3, the Overpass Turbo is able to provide an

accurate, up-to-date and handy access to almost all major cities across the world.

We can locate a target area by running a query to search the name of that city

or region, and zoom in or out to an appropriate size. By dragging a rectangular box

to select the zone of interest, the buildings, structures and trees within the zone are

highlighted. Notice that the absolute coordinates and heights of all objects in the zone

are contained. We can then extract these data into a portable file, e.g., kml format,

which incorporates global longitudes and latitudes of the structure vertices as {lon, lat}pairs.

The bottom left corner of the drag box should be specially marked as it will be used

as the Origin in our ray launching simulator. If the global longitude and latitude of

29

this point are (Ox, Oy), the Earth radius, r, at this latitude can be derived from Eqn.

(3.2);

r =

√√√√ [r2max cosOy]2 +

[r2min sinOy

]2[rmax cosOy]

2 + [rmin sinOy]2 (3.2)

given the earth maximum radius at the equator, rmax =6378.137km, and minimum

radius at either of the poles, rmin =6356.752km. Subsequently, the relative coordinates

of a point, P (px, py), can be obtained via Eqn. (3.3);

(px, py) = [sin (loni)−Ox, sin (lati)−Oy] · r (3.3)

Following the above procedures, the structures on the map are transformed into

polygons on a familiar Cartesian xy-coordinate system with the Origin located at

the bottom left corner. The flat surfaces can then be extracted for ray intersection

detection.

Figure 3-3: The Arc de Triomphe area, Paris, on the Overpass Turbo panel.

An example of the Arc de Triomphe area in Paris is shown in Figure 3-3. The

highlighted buildings and structures are marked with a blue outline with bold yellow

facets. Trees and smaller segments are covered by purple circles. The height of each

structure is also stored in the file.

30

The extracted urban layout can be plotted in Matlab as displayed in Figure 3-

4. A transmitter is added on the Champs Elysees Avenue marked as a red dot. A

receiver is located on a nearby road. A series of rays are radiating uniformly from the

transmitter marked as blue lines, which are subsequently reflected after intersection

with flat surfaces. Obviously, the black polygons represent the buildings as viewed

from above. The green circles are the trees. There are two red lines, one connecting a

corner of the Arc de Triomphe and the transmitter, the other joining the same corner

with the receiver. This is one of the diffraction paths being detected. A light blue line

goes across the Arc de Triomphe to arrive at the receiver, which is a wall penetration

instance. The diffraction and wall penetration are always tracing for each ray, but will

only be displayed if the out-going ray reaches the receiver.

Note that only a limited number of rays are shot from the transmitter, which results

in quite a few missing detections of apparent edge diffractions and transmissions. The

coverage range of wave propagation is also restricted with areas of blind spot. This

is primarily due to an insufficient number of rays being launched. In practical simu-

lations, less than 0.5◦ of angular increment is usually allocated between neighbouring

ray departures which generate over 720 rays. More densely shot rays guarantee more

detailed intersection detections. However, as the number of rays increases the compu-

tation load also increases. Consequently, more comprehensive ray arrival information

can be obtained at the receiver.

Figure 3-4: Propagation paths calculated using Matlab for the Arc de Triomphe area,Paris, given a virtual RF source at (630m, 100m). All reflected rays are presented inblue, whereas, only penetrated, diffracted and root-top diffracted rays that are receivedare shown in light blue, red and orange, respectively.

31

This simulation shown in Figure 3-4 at the Arc de Triomphe area takes over 1000

s, provided that only one transceiver channel is evaluated. Methods used to accelerate

the simulator are detailed in section 6.

On the other hand, overly dense rays being shot means that double counting of

rays is likely. When neighbouring rays going through an identical propagation path

are received, all rays but one should be omitted since they represent a same wavefront.

This selection function has been implemented by identifying rays that depart from a

similar angle while going through the exact same sequence of surfaces and end up with

similar travelled distances.

3.2.3 Trees

The modelling of wave propagation through vegetation can be very complicated.

Firstly, in urban environments there are thousands of different kinds of trees with var-

ious shapes, heights and foliage densities. These properties usually change seasonally.

The random layout of leaves and branches introduces complex diffractions, transmis-

sions and scattering [38], which prevents a both accurate and efficient model.

Among a large number of studies on measurements and modelling of trees [96, 98,

114]. Benzair et al. provided a reliable empirical path loss model on wave propagation

through trees [13]. This model considers the frequency (from 1-4GHz), interception

distance across the tree crown, as well as the season. The path loss formula is provided

in Eqn. (3.4);

Ltree = a× f b × df (3.4)

in which f is the frequency in GHz, df is the tree depth traversed by the ray, a and

b are constants from the lookup Table 3.1, which vary depending on seasonal foliage

density as well as humidity.

Table 3.1: Table on parameters a and b of the tree loss model

Season Summer Winter

Humidity Median 50% 90% Median 50% 90%

a 0.57 0.71 0.78 0.36 0.52 0.59

b 0.60 0.47 0.42 0.43 0.29 0.25

We use this model to predict the path loss through trees. The scattered rays by the

tree are neglected since the power of these rays is so small that tracing them adds much

computational burden but adds little to the accuracy [98]. In our 2.5D ray launching

32

model, these trees of which the precise locations can be accessed from the OSM are

always treated as a cylindrical attenuating media with 3m radius. Hence, rays going

through trees will be subject to power loss but do not change direction. An example

of the transmission through a tree is shown in Figure 3-5.

Figure 3-5: Transmission of a ray through a tree in which the crown is treated as acylindrical attenuating medium.

3.3 Propagation primitives

Four propagation primitives are considered in the ray launching model including

reflection, edge diffraction, wall penetration and roof-top diffraction. It is believed these

propagation phenomena make the most significant difference to channel estimation [11].

These models are integrated in a variety of highly recognized models and measurements

found in the literature. The mechanisms and empirical models in these primitives are

explained in detail in this section.

33

3.3.1 Reflection

In urban environments, reflection plays one of the most significant roles which gives

rise to multipath propagation. If the LOS path is blocked, the physical channel between

a transmitter and a receiver is usually made of reflections from flat building surfaces.

Studies of corridor models show typical propagation environments in the street canyons

when both the transmitter and receiver are below the roof-top height [52]. Bertoni and

Walfisch-Ikegami models [105], in which the signals are diffracted at the roof edges,

also investigate the reflection phenomenon of rays in the vertical plane. In these cases,

it is critical to determine the reflection mechanism of microwaves for building surfaces.

The frequency range of modern communication systems, e.g., radio broadcasts,

DAB radio, GSM and Wi-Fi, is usually between 1MHz and 10GHz [94, 111]. The

wavelengths of such signals are greater than 3cm. The modern building surfaces are

mostly made of concrete, glasses, aluminium alloys or plastics, of which the surface

roughness can be neglected compared to the wavelengths. Therefore, rough surface

scattering is not taken into account. Hence, specular reflection generally applies to

microwave intersections with building surfaces [67]. However, the polarization of the

microwave after reflection varies depending on the surface material as well as the inci-

dence angle.

According to the UTD, a vertically polarized wave generates both vertically polar-

ized and horizontally polarized waves after reflection as long as the incidence angle is

within 0◦ to 90◦ [95]. Note that the incidence angle here indicates the angle between

the projection ray and the flat surface. However, the vertically polarized wave com-

prises most of the reflected energy as stated in [75, 86]. In the 2.5D ray launching

model, we assume that rays launched from the transmitter are vertically polarized, and

horizontally polarized waves after reflection can be neglected.

Specular reflection dictates that the reflection angle is always equal to the inci-

dence angle within the intersection plane so that the direction of reflected ray can be

determined. A reflection situation is demonstrated in Figure 3-6. A scalar, called the

reflection coefficient, is used to describe how much electromagnetic energy is dispersed

in that ray with respect to incidence angle and surface material [76]. According to the

assumptions of our model, the reflected ray of a vertically polarized incident wave is

also vertically polarized, and the reflection coefficient as follows Eqn. (3.5);

Rvertical =ε · sin θ −

√ε− cos2 θ

ε · sin θ +√ε− cos2 θ

(3.5)

where θ is the incidence angle, ε represents the electric permittivity of the surface

material which can be obtained from Eqn. (3.6);

34

ε = εr − j60σλ (3.6)

where εr is the relative electric permittivity, σ indicates the electrical conductivity, λ

equals the wavelength.

Figure 3-6: Reflection geometry on flat surface.

It is apparant that the electrical properties of the surface materials should be known

as a priori. Provided the locations of buildings from digital map databases, we can

assign surface material properties to each structure depending on the data available or

simply assuming a common material. For instance, the relative dielectric constant and

conductivity of concrete are 4.0 and 10−4 S/m [93], respectively. They can be assigned

to an unknown surface if corresponding information is not available since concrete is

typically the most common surface material in urban environments.

Once a propagation ray intersects the surface, the reflected direction and signal

strength of the reflected ray can be determined. The reflected ray makes a new virtual

source and can continue to be traced onwards until it satisfies one of the fading-out

criteria, including shooting out of the zone of interest, getting reflected for more than

seven times, or attenuating below -140dBm. Notice that convergence tests suggest that

rays reflected more than seven times make little difference to the channel estimation

[12]. If a ray is reflected over seven times it is regarded as terminated. Furthermore, it

is difficult for mobile measurement devices to detect signals below -140dBm.

35

3.3.2 Diffraction

Diffraction occurs when a ray hits the edge or cusp of a structure. In densely built-

up environments, the vertical edges of buildings act as important media for RF signals

to propagate along street canyons. In the Bertoni and Walfisch-Ikegami models [105],

the horizontal roof edges illuminate the NLOS areas between rows of buildings. The

tip of an object may also give rise to a spherical diffraction.

The GTD is regarded as a benchmark interpretation of the diffraction mechanism

[54]. Depending on different incidence angles of a ray, it could generate a cone-like

wavefront or a 2D circular wavefront normal to the edge [60]. A cusp of an object

basically disperses the ray in all directions. The power of the diffracted ray is a function

of several factors including arrival angle, departure angle, the distance from the source

to the edge, the distance from edge to the receiver and reflection coefficients of the two

surfaces that form the edge.

According to the literature on the validation of diffraction [9, 85], one order of

diffraction is sufficient in urban environments due to the huge computational effort

using higher order diffractions for only a limited increase in accuracy. For our 2.5D

ray launching model, we assume the vertical edge and cusp of structures become new

virtual sources and disperse the ray in all directions. Similar to reflection, we define a

scalar, called diffractivity, to measure the power ratio of the out-going ray compared

to the incident ray.

Figure 3-7: Diffraction geometry on the edge of a wedge.

36

Figure 3-7 shows a ray diffracted at the edge of a wedge in which n is the field angle.

Thus, the interior angle can be expressed as (2 − n)π. Term ϕ is the incidence angle

with respect to the incidence wedge facet, and ϕ′ is the departing angle with respect

to the same facet. According to [11], diffractivity can be obtained from Eqn. (3.7);

D =−e−j(π/4)

2n√

2πβ· { cot

(π + (ϕ− ϕ′)

2n

)· F[βLa+(ϕ− ϕ′)

]+ cot

(π − (ϕ− ϕ′)

2n

)· F[βLa−(ϕ− ϕ′)

]+R0 · cot

(π − (ϕ+ ϕ′)

2n

)· F[βLa−(ϕ+ ϕ′)

]+Rn · cot

(π + (ϕ+ ϕ′)

2n

)· F[βLa+(ϕ+ ϕ′)

]}

(3.7)

where term β represents the wave number, R0, Rn refer to the reflection coefficients

of the incidence wedge facet and opposite wedge facet. The reflection coefficients can

be calculated using Eqn. (3.5) with respect to the incident wedge facet and opposite

wedge facet, respectively. The Fresnel integral F (x) is given by;

F (x) = 2j√xejx

∫ ∞√xe−jγ

2dγ (3.8)

and L is a distance term in 2D scenarios defined as;

L =s · s′

s+ s′(3.9)

where term s′ denotes the distance from source to the diffraction edge, and s is the

distance from the edge to the receiver. The angle terms a± are given by;

a±(ϕ± ϕ′) = 2 cos

[2πnN± − (ϕ± ϕ′)

2

]2(3.10)

in which the integers N± are those which most closely satisfy Eqn. (3.11) and Eqn.

(3.12);

2πnN+ − (ϕ± ϕ′) = π (3.11)

2πnN− − (ϕ± ϕ′) = −π (3.12)

37

3.3.3 Wall penetration

The concrete walls and glasses may also allow penetration of microwaves. Many

measurement and modelling attempts have been made to describe RF transmission

mechanisms through walls. Ding et al. [69] carried out wall-penetration measurements

at 1.93GHz. Okvist et al. [89] carried out measurements at 15GHz for 5G application.

Pena et al. [90] worked on 900MHz wall-penetration tests and developed a statistical

model. Jiang et al. [50] put forward a specific air-to-concrete transmission model.

Deterministic models usually take into account the thickness of a wall and determine a

refraction path within the wall. Multiple reflections and transmissions within the wall

layer can also be elaborated as shown in Figure 3-8.

Figure 3-8: Comprehensive transmission geometry for modelling wall penetration.

However, the computational effort of many of these models is massive, and the

directional shift due to refraction is often very small in urban scenarios. Moreover, the

information on material and thickness which can be difficult to obtain and are essential

to build such models.

38

The researchers in [111] have put forward a semi-deterministic model which ap-

proximates losses through one layer of wall regardless of the thickness. This model

considers the frequency of the signal and its incidence angle which prove to be the

most significant variables regarding the penetration loss.

The loss with respect to the frequency of the incident signal is provided in Eqn.

(3.13);

Lconcrete,dB = 4fGHz + 5 (3.13)

where fGHz is the frequency in GHz (in the range of 0.1-60GHz). The loss as a function

of incidence angle is derived from Eqn. (3.14);

Langle,dB = 20× (1− cos θ)2 (3.14)

where the incidence angle is denoted as θ.

The loss is presented in dB units. Hence, the signal strength drops by Lconcrete,dB+

Langle,dB after each penetration without changing direction. In order to integrate this

model into our ray launching engine, the number of transmission is recorded for each

ray along its propagation path, and the losses will be deducted from the RSS during

field reconstruction phase.

Up to three layers of wall penetration are allowed as vast attenuation is estimated

for more than that. Note that reflection on the indoor surfaces for the penetrated ray

is also a concern because in many cases the inner surface makes a multipath component

before the rays arrive at the receiver. Since the indoor configuration can be too complex

to model and the rays may find unexpected scatterings, only one order of inner surface

reflection is taken into account where the order of penetration before or after indoor

reflections is also restricted to one.

3.3.4 Roof-top diffraction

As the scale of the propagation domain increases, reflections, vertical edge diffrac-

tion and wall penetrations below the roof level become less detectable [59]. Besides

the free space propagation loss, higher orders of these propagation behaviours result in

the power degrading exponentially. The roof-top diffraction becomes dominant in such

long distance wave propagation paths.

The propagation over roof tops is indeed very common in urban environments. For

example, the LOS path from a base station to our mobile communication devices is

always blocked in high-rise areas. A street view of New York city is shown in Figure

3-9. Thanks to the horizontal edges of buildings, the signal is able to perform multiple

39

forward diffractions past rows of buildings and go down from the nearest roof edge to

the street [80]. This process can be divided into three stages, from the emitter to the

first diffraction edge, multiple diffractions past building tops and from the last roof

edge to the mobile.

Many researchers in the literature have shown measurement results at different fre-

quencies in different cities, and proposed some empirical and semi-deterministic models

to estimate the path loss [65, 81, 120]. Xia et al. presented an analytical model [134]

which categorized such propagation behaviours into three circumstances, when the

emitter is: above the roof-top height in Eqn. (3.24); about the roof-top height in Eqn.

(3.25); below the roof-top height in Eqn. (3.26). It considers the rows of building as

multiple knife-edge screens [27, 47, 121]. A diagram illustrating a below roof-top case

can be seen in Figure 3-10.

Figure 3-9: New York street canyon view taken from:www.ecology.com/2013/01/03/green-walls-cut-street-canyon

The model in [134] considers a variety of factors such as heights of the transmitter

and receiver, distance from transmitter to the first diffraction edge, distance from

the nearest roof edge to the receiver, travelled length horizontally, average separation

distance between rows of buildings etc. This model is relatively accurate, and at the

40

same time, computationally efficient. The roof-top diffraction is mainly based upon

this model.

Figure 3-10: The geometry of roof top diffraction when the transmitter height is belowaverage building height.

The propagation loss in the last stage, where rays drop from roof edge to the street,

is given by Eqn. (3.15);

Lrtd = −10 log

[λ

2π2r

(1

θ− 1

2π + θ

)2]

(3.15)

where λ is the wavelength. The receiving angle in the vertical plane, θ, and the distance

from the edge to the receiver, r, can be expressed as follows;

θ = tan−1(

∆hmx

)(3.16)

r =√

(∆hm)2 + x2 (3.17)

where ∆hm is the height difference between the nearest roof edge and the receiver, and

x is their horizontal distance.

The multiple screen diffraction loss can be expressed as;

Lmd = −10 logQ2M (3.18)

where variable QM is given by;

QM =√M

∞∑q=0

1

q!(2g√jπ)qIM−1,q (3.19)

where M is the number of diffracting screens. Given in Eqn. (3.20), term g is defined

as;

41

g = ∆hb1√λd

(3.20)

where ∆hb denotes the height difference between the transmitter and the first edge,

and d indicates the average separation distance between rows of buildings.

In Eqn. (3.19) IM−1,q satisfies a recursion relation as follows;

IM−1,q =(M − 1)(q − 1)

2MIM−1,q +

1

2√πM

M−2∑n=1

In,q−1√M − 1− n

(3.21)

with initial terms equal to;

IM−1,0 =1

M2/3(3.22)

IM−1,1 =1

4√π

M−1∑n=0

1

n2/3(M − n)2/3(3.23)

The path loss derivation can be simplified using a recursive approach considering

building heights are uniform [134]. Given the total travelled distance from transmitter

to receiver, R, the formulas used to determine propagation loss according to transmitter

height at roof-top level, above roof-top level and below roof-top level are listed in Eqn.

(3.24), Eqn. (3.25) and Eqn. (3.26), respectively;

Lat = −10 log

{(λ

2√

2πR

)2

·

[λ

2π2r

(1

θ− 1

2π + θ

)2]·(d

R

)2}

(3.24)

Labv = −10 log

(

λ

4πR

)2

·

[λ

2π2r

(1

θ− 1

2π + θ

)2]· (2.35)2 ·

(∆hbR

√d

λ

)1.8(3.25)

Lbl = −10 log{(

λ

2√

2πR

)2

·

[λ

2π2r

(1

θ− 1

2π + θ

)2]·[

d

2π(R− d)

]2· λ√

(∆hb)2 + d2·(

1

φ− 1

2π + φ

)2

}(3.26)

We integrate these models in our ray launching engine. In a 2D horizontal domain,

a straight line connecting directly from transmitter to receiver is regarded as the prop-

agation path. If the number of intersections with surfaces along this path is K, the

42

integer of knife edges approximates (K − 1)/2, which will be used as M with regard to

the model (Eqn. (3.19)). In vertical domain, the total travelled distance is comprised

of the distance from transmitter to the first edge, and then to the last edge, then to

the receiver. The average separation distance between buildings can be easily obtained

by dividing the horizontal distance by M . Depending on the height of transmitter

compared to average roof-top height in the area, one of the formulas will be adopted

to calculate the path loss.

3.4 Field reconstruction

The ray launching models rely on a geometrical capturing mechanism to decide

whether a ray is truly arriving at the receiver or not. Using a reception sphere is a

commonly acknowledged method in the literature [66, 131].

The propagation paths in a specific environment are solely determined by the trans-

mitter location and the object layout. In other words, we assume that the receiving

antennas, including users, are making little difference to electromagnetic fields.

Figure 3-11: An illustration of the reception detection mechanism in which a virtualreception sphere is built at the receiver.

A reception sphere is a virtual circle centered at the receiver. The radius of the

sphere is specified depending on the density of rays in the scenario. A search of the

literature suggests that the scale of a reception sphere should be appropriate to only

accommodate one ray at a time [100] as shown in Figure 3-11. This means that the

radius of a sphere varies from location to location as neighbouring rays diverge when

43

they depart farther from the transmitter. In our model, a uniform radius is set in which

double counted neighbouring rays are discriminated and treated as one capture.

Multipath propagation is very common in urban scenarios, which can be realized

and analysed by wave propagation models. A reception sphere regards any propagating

ray that falls within the circle as being received. The channel estimation is implemented

by superposition of all rays that are deemed to arrive.

A single arrival may come from a combination of reflections, vertical edge diffrac-

tions, transmissions or directly from the nearest roof top. Three characteristics that

are of our main interest are the arriving angle, delay and power. The arrival angle can

be easily readable from the capturing geometry at the reception sphere.

The time delay, TOA, is approximately the total travelled distance, S, which should

be carefully recorded, divided by the speed of light, c. The derivation of time delay

can be written as;

TOA =S

c

With respect to [6, 11], the expression for computing the received power can be

written as Eqn. (3.27);

Pr = Pt ·GtGrλ

2

(4πd)2

∏j

Rj

2 ∏j

Dk

2

− 10−LT /10 − 10−Lrtd/10 (3.27)

where Gt, Gr are the transmitting and receiving antennas gains, respectively, Rj rep-

resents the reflection coefficient of the j th intersecting surface, and Dk refers to the

diffractivity of the kth diffracting edge. Terms LT and Lrtd are transmission loss and

roof-top diffraction loss, respectively.

The received characteristics allow us to determine the Power Delay Profile [57, 138].

A PDP is able to exhibit the received power of each ray versus the arriving time

individually. It collectively depicts the channel impulse response.

The ray launching model is also capable of identifying directions of arrival with

respect to the power. A polar coordinate plot, also called Angular Spread, at the

receiver is used to specify the AOAs [23, 61, 116]. These arrivals can also be traced

back to the source where rays departing from the very first wavefront are recognized.

A polar coordinate plot at the transmitter shows in which directions of departures

arriving detections are made at the receiver. Lastly, the radiation patterns of the

antennas should be taken into account (Figure 3-2).

A virtual environment consisting of an enclosed box is designed to demonstrate how

field reconstruction works (Figure 3-12). Seven objects of arbitrary shape are put in

44

an enclosed box which is 120m×120m square. Given the transmitter marked as a red

dot, the received is a yellow dot positioned in the LOS area. The blues segments are

either rays being shot or their reflections within seven orders of reflection. The magenta

arrows refer to those rays fading out after seven reflections. Several red lines mark the

diffraction paths. Wall penetration cases that are being received are indicated with

light blue lines. Roof-top diffraction does not apply to this case.

Figure 3-12: Propagation paths of an imaginary emitter at (33m, 47m), based on avirtual enclosed box scenario with seven arbitrary objects inside. All reflected rays arepresented in blue, whereas, only penetrated and diffracted rays that are received areshown in light blue and red, respectively.

With the help of the reception detection function, receiving information can be

gathered. Correspondingly, a PDP and two Angular Spreads are displayed in Figure 3-

13 and Figure 3-14. Notice that the impulses in PDP are coloured with respect to their

types of propagation. The earliest arrival is obviously the LOS ray with the largest

signal strength, at approximately -75dBm. The very last arrival takes almost 0.85ms,

which suggests a total travelled distance around 255m within the enclosed box.

The spikes in the Angular Spreads are labelled in dB relative to the minimum signal

strength detected. It can be easily seen that most arrivals are clustered within a 180◦

sector of the receiver, which is intuitively recognizable given the geometry. On the

45

other hand, those arrivals may come from any direction from the original wavefront at

the source.

Figure 3-13: Power Delay Profile of the receiver at (90m, 60m) in the seven-objectscenario.

Figure 3-14: Angle of Departure and Arrival from the perspectives of transmitter andreceiver, respectively, in the seven-object scenario. The receiver is located at (90m, 60m).

46

Generally, the PDP and the Angular Spreads present a realistic estimation of the

received power, time delay of arrivals and angle of arrivals. Given the environment

settings, the ray launching model successfully estimates electromagnetic features with

respect to the transmitter and receiver locations. In this work, we consider how those

features can be exploited to aid navigation and source localization.

3.5 2.5D ray launching model validation

Figure 3-15: Measurement campaigns in Munich downtown area with three routes,‘METRO 200’, ‘METRO 201’ and ‘METRO 202’.

47

In order to validate our 2.5D ray launching model, we ran a full simulation of

the Munich downtown area where channel sounding measurements at 947MHz along

three different routes are accessible. This data has been created during the COST 231

action as described by Damosso (1999) and is now publicly available at Mannesmann

Mobilfunk GmbH, Germany [36, 46]. Figure 3-15 shows the actual layout of the area

as well as the emitter location and three routes shown in different colours. Further

analysis of this scenario can be found in Chapter 5.

The scale of this scenario is 2.4km× 3.4km and contains 2088 buildings, i.e., 17455

walls. Also within this area are 1758 trees accessible from the Overpass Turbo database.

The emitter was located at (1281.36m, 1381.27m) and 13m above the ground, while

the receiver was being held at a height of 1.5m in all three cases.

Discrete measurements of path loss were made along these routes with nearly the

same distance interval. Omni-directional antennas were used during the measurements.

A total of 970, 355 and 1031 observations were recorded along the three routes, respec-

tively. The trends of path loss variation against sampling sequences are shown in Figure

3-16.

Figure 3-16: Pathloss trends along the three measurement routes on the Munich sce-nario.

For the simulation, the Origin point during measurement was found and aligned

48

with that in the model. A transmitted power equalled to 1W with omni-directional

radiation pattern was established. After setting up buildings and trees, an average

building height of 18.7m was adopted for roof-top diffraction loss analysis. Due to

incomplete data on building materials, we assumed they were all concrete regardless

of thickness. Reflection and diffraction limits were set as seven and one, respectively.

Three orders of wall penetration and one order of indoor reflection were considered.

Reception spheres of 5m radius which inscribes the grid of pixels was allocated across

the entire scenario. The RSS for each sampling position was taken from the nearest

reception sphere. As a result, we have connected those discrete RSSs into data series

to obtain path losses. The path losses are plotted against the original measured data

along the three routes as shown in Figures 3-17, 3-18 and 3-19.

Figure 3-17: Simulated RSSs along ‘METRO 200’ compared to measurements.

49



50

Generally, the simulated path loss variations show a remarkable agreement with

the measurements. The major trends of fluctuations which are presumably due to the

mobile device threading through LOS and NLOS areas have been successfully depicted.

The path loss predications show a good agreement with the measured data. This is

true for both the close LOS areas as well as for the more distant NLOS areas where

roof-top diffraction is important.

Table 3.2 shows the Mean Absolute Errors and Standard Deviations between simu-

lation results and measurements along the three routes. As can be seen, the Standard

Deviation always remains less than 7dB along any route. The Mean Absolute Error

also shows good agreement. Overall, the ray launching model is shown to be effective

for a relatively large and complicated urban scenario.

Table 3.2: Mean Absolute Errors and Standard Deviations along the three routes inthe Munich scenario

Route ‘METRO 200’ ‘METRO 201’ ‘METRO 202’

Mean Absolute Error(dB) 7.7373 5.7189 6.4210

Standard Deviation(dB) 6.6026 6.4883 6.8691

Nevertheless, disagreements can be found. Several factors may potentially account

for these errors. Firstly, the environmental entry does not provide complete details,

e.g., the precise building geometries, heights, and materials, which will certainly restrict

the accuracy. Secondly, the 2.5D model has its limitations in dealing with certain prop-

agation primitives, such as ground reflections, or higher orders of diffraction involving

both horizontal and vertical edges. Thirdly, the presence of vehicles and pedestrians

are very much likely to affect the propagation circumstances in practice.

Taking the longest measurement route ‘METRO 200’ for example, we plotted dis-

crete simulated path losses against measurements as shown in Figure 3-20. This scatter

plot leaves out the spatial consecutiveness of measurements while regarding each path

loss value as an individual event. The linear correlation coefficient along ‘METRO

202’ equals to 0.9089 which demonstrates good agreement. Generally, the simulations

and measurements are linearly dependent which verifies the effectiveness of our ray

launching model. However, we do notice that in the smaller path loss range the sim-

ulator tends to underestimate the actual measured loss. This is most likely due to

limitations in incorporating all environmental information, e.g., traffic and pedestrians

which prevent LOS propagation. The routes ‘METRO 200’ and ‘METRO 201’ produce

correlation coefficients of 0.9236 and 0.8725, respectively. Thus, all routes show good

agreement between simulation and measurement.

51

Figure 3-20: Scatter plot between simulated pathloss and measured pathloss along‘METRO202’. The solid blue line shows the trajectory of correlation coefficient equalledto unity.

Wave propagation modelling aims at producing as accurate channel estimations as

possible. However, even the best current commercial ray tracers are not able to achieve

perfection. The core of this research lies in the leveraging of coarse but efficient channel

predictions to accomplish navigation and source localization. These are to be discussed

in the following chapters.

3.6 Summary

This chapter expands the design procedures of the 2.5D ray launching model. This

model integrates geometric information of the environment along with the most signif-

icant propagation mechanisms of free-space propagation model. Four principal propa-

gation mechanisms are incorporated, including reflection, diffraction, wall penetration

and roof-top diffraction. A virtual reception sphere is used to detect arriving rays.

This model has been validated against measurement campaign in Munich city center

demonstrating very good agreement.

52

Chapter 4

Navigation

Many modern navigation systems are heavily reliant upon the GNSS. Whereas,

densely built urban environments which are vulnerable to multipath propagations un-

dermine the precision and reliability. There is currently growing interest in localiza-

tion based on signals of opportunities in which propagation parameters observed from

ground emitters may allow urban navigation. Although these methods take advantage

of the multipath, they usually require laborious collections of ‘truth data’ at various

locations. Furthermore the temporal stability of the environment often reduces the

effectiveness.

In this chapter, we propose to obtain channel parameters from modelling instead

of field measurement, and exploit a fingerprinting approach to bond those parameters

to locations. At the heart of this method lies the definition of the ‘location fingerprint’

and the development of a reliable localization algorithm. We demonstrate what com-

ponents are essential in constituting the fingerprint and how to extract them from our

2.5D ray launching model. It is also found that Artificial Neural Networks are a valid

tool to generate the mapping functions. In pursuit of an optimized mapping function,

we investigate a series of variables and associated errors affecting the on-line naviga-

tion process. Overall, the proposed navigation system shows very good accuracy and

reliability.

4.1 Location fingerprint

Much research in the literature has validated the feasibility of using field measure-

ments to locate a mobile device. The most popular primitive to exploit is RSS because

the embedded device is easy to install and there are plenty of wireless communication

access points. RSS measurements are used for indoor localization systems in [118, 132].

53

Abid et al. [4] experimented and validated using RSS measurements to navigate in ur-

ban areas. Wilfinger et al. [127] proposed a position determination algorithm based on

RSS readings from passive RFID (Radio Frequency Identification Devices) tags. The

digital television, radio cellular base station, Wi-Fi source etc. are common places in

urban environments nowadays. As a result, signals of opportunities are rich resources

to take advantage of.

Although conventional triangulation-based localization methods rely very much

upon TDOA and AOA measurements, their reliabilities drop significantly in severe mul-

tipath environments [101]. Despite some disadvantages, the capability of using TDOA

and AOA to navigate has recently drawn considerable research attention [22, 110] as

long as assuming LOS-only propagation.

In the literature review, we have addressed that spatially varying feature is per-

vasively seen in the RSS, TDOA and AOA distributions of these signals. Therefore,

the core concept of our navigation system is utilizing the spatially-distributed channel

characteristics to discriminate individual physical locations. Even though the multi-

path breaks the elegance of simple dependence in terms of power attenuation, delays

on travelled distance and angular spread, it brings about diversity and irregularity into

the data distribution which gives rise to viability of location discrimination.

An advantage of propagation modelling is that it generates channel response es-

timations on top of a balance between accuracy and rapidity which can be balanced.

Given a description of the environment, we are able to predict receiver observations

rather than carrying out laborious field measurements. Even if a detailed geometry is

to be simulated, it is likely to take much less than measurement. Moreover, changes in

the environment, e.g., construction of a new building, adding interfering emitters and

the presence of traffic, are merely costing another round of simulation. Whereas, such

changes could lead all measurement efforts to be in vain.

An essential step before establishing a navigation system is to extract geographical

features from channel characteristics. By way of example, a virtual enclosed box is built

in the last chapter. The same geometry is used with a receiver located in a NLOS area

as is shown in Figure 4-1. Subsequently, the PDP and Angular Spread are obtained

and displayed in Figure 4-2 and Figure 4-3. It can seen that the features extracted

from this location are significantly different from those in Figures 3-13 and 3-14. Our

hypothesis is that the characteristics are hopefully distinctive in comparison to any

other receiver location. To solve the navigation problem, we need to identify a method

of mapping the received characteristics to a location. A popular technique, known as

fingerprinting, is adopted here to solve this data mining issue.

54

Figure 4-1: Propagation paths to a NLOS receiver at (114m, 32m) on the virtual seven-object scenario. All reflected rays are presented in blue, whereas, only penetrated,diffracted and roof-top diffracted rays that are received are shown in light blue, redand orange, respectively.

In this thesis, a fingerprint which represents a physical location is termed as location

fingerprint. Ideally, a location fingerprint must be unique. Taking received power for

example, an omni-directional antenna may find similar readings in completely opposite

directions to the emitter. Thus, features of limited dimensions alone usually cannot

make an effective location fingerprint. Previous fingerprinting approaches, e.g., [113],

that absolutely rely on RSS measurements do not produce good accuracy performances

with less than 20 signals of opportunities. Although more components in a location

fingerprint usually guarantees better uniqueness, a-priori knowledge of many sources

becomes critical. In this research we aim to reducing this number to less than ten by

incorporating TDOAs and AOAs as well as enhancing the localization algorithm.

55

Figure 4-2: PDP of the NLOS receiver at (114m, 32m).

Figure 4-3: Angular Spread of the NLOS receiver at (114m, 32m).

A location fingerprint primitive which is the element of a location fingerprint should

be repeatable. This means that different users should expect a similar reading using

56

an identical device at the same location. Only when this expectation holds could a lo-

calization algorithm work. Provided good accuracy capability from the ray launching

model as well as easily measurable electromagnetic characteristics, it ought to be pos-

sible to create a suitable location fingerprint primitive [140]. These primitives must be

relatively stable in an ever changing electromagnetic environment (e.g., due to pedes-

trians and vehicles), such that training and deep learning result in reliable and resilient

navigation solutions.

From the PDP and Angular Spread, a series of eigenvectors with fingerprinting

significance can be extracted, such as RSS, MED (Mean Excess delay), RMSDS (Root-

Mean-Square Delay Spread) and AOA (Angle of Arrival). These eigenvectors all possess

a degree of uniqueness and repeatability which form good location fingerprints.

Given the received power, Pr, of each arriving ray, the average received signal

strength, RSS, at location (x, y) can be approximated as Eqn. (4.1);

RSS(x,y) =

√√√√ N∑i=1

Pri2 (4.1)

where N is the total number of impulses. The MED at (x, y) is given by;

MED(x,y) =

∑Ni=1 Priti∑Ni=1 Pri

(4.2)

where ti is the absolute time delay of each arriving ray. The RMSDS at (x, y) can be

expressed as;

RMSDS(x,y) =√τ2 +MED2

x,y (4.3)

where τ2 is derived from;

τ2 =

∑Ni=1 Prit

2i∑N

i=1 Pri(4.4)

In this thesis, we assume only the most significant AOAs can be measured and only

with low accuracy. Hence, we suppose the first AOA is exactly the direction in which

the strongest received power is detected. The second AOA refers to the direction of

second strongest ray and so on. A direction is a vector which has to be presented with

at least two scalars. Thus, we usually represent the AOA as Cartesian projections;

−−−→AOA =

(cos−−−→AOA, sin

−−−→AOA

)= (AOA1X,AOA1Y)

where AOA1X refers to the x -axis Cartesian projection of the most significant arrival,

57

and AOA1Y refers to the y-axis Cartesian projection of the most significant arrival.

The projected scalars of an AOA are always between [-1, 1]. Similarly, we would

like to normalize the RSS, MED and RMSDS within an equivalent range depending on

the measurement requirements. For example, RSSs are truncated between [-140dBm,

-60dBm] because below -140dBm signals are hardly detectable and signals in the far

field are generally below -60dB.

Ultimately, we aim to integrate these features by merging the normalized values to-

gether to construct a location fingerprint. Given M signals of opportunities, a location

fingerprint as a function of 2D coordinates can be written as

Fingerprint(x, y) =

RSS1 MED1 RMSDS1 AOA1X1 AOA1Y1



......

......

...

RSSM MEDM RMSDSM AOA1XM AOA1YM

Figure 4-4: A schematic diagram on fingerprints generated by opportunistic sources.

In the physical layer, each known source of opportunity would generate a row of

fingerprint primitives regarding each location as illustrated in Figure 4-4. Each column

stands for an extracted and normalized electromagnetic feature for a location. A loca-

58

tion can be redefined by tagging such fingerprint matrix, and we would like to develop

a localization algorithm using these tags.

4.2 Localization algorithm

The location fingerprint is obtained to represent the identity of each geometric lo-

cation which incorporates RSS, MED, RMSDS and AOA. Then the fingerprint can

be further expanded by considering contributions of opportunistic signals, e.g., digital

television, radio broadcasts, cellular communications, Wi-Fi signals etc. In this way,

the dimension of the location fingerprint is magnified by the number of opportunistic

sources in the scenario. Since the source locations are usually different in practice,

their propagation geometries differ from each other. On the other hand, as the prop-

agation coefficients such as the reflectivity of surfaces are frequency dependent, the

receiving eigenvectors will vary to allow diversity for the fingerprints. Overall, the

more signals available in the scenario the richer each fingerprint will be. Modern urban

areas generally satisfy this condition as a variety of transmitters are sited for different

telecommunications purposes.

The localisation algorithm is developed based on the idea of exhaustive searching

and matching of the location fingerprints. Given the known locations of sources in

the area of interest, we run the simulation for all candidate locations of the receiver.

These location fingerprints are subsequently stored in a database. Suppose a mobile

user wanders through the given scenario along an unknown path while continuously

reporting their measured fingerprints. The system will find the best match of loca-

tion fingerprint within the database to decide their location. Note that for the same

frequency we assume there is only one source such that the sources do not interfere.

4.2.1 Mapping function

A fingerprint-based localization algorithm is generally comprised of on-line phase

and off-line phase. During the on-line phase, when an entry of observed fingerprint is

given the algorithm is expected to return a best matched location among the fingerprint

database. Provided that we have successfully defined and extracted fingerprints in the

off-line phase, a reliable mapping function bridging the fingerprints and the locations

becomes the key.

Apparently, the location in a 2D coordinate system can be expressed as (x, y) of

which the absolute values are related to the origin. On the other hand, the finger-

prints have been defined as matrix comprised of electromagnetic features including

RSS, MED, RMSDS and AOA. We attempted to visualize the fingerprint primitives by

59

allocating a grid of receivers of 1m×1m square and plotting these primitives against

the coordinates. As can be seen from Figure 4-5, a heat map of RSS is built for the

scenario where brighter coloured pixels imply higher RSS and darker pixels indicate

lower RSS detected. Not only are the shading effects of objects presented, but also the

radiation pattern contours of the emitter in the horizontal domain are also reflected as

provided in Figure 3-2.

Figure 4-5: RSS heat map in the enclosed box scenario with seven objects given anemitter transmitting 0dBm at (33m, 47m).

A significant characteristic of the RSS data distribution is their discontinuous na-

ture. Although the pixels turn darker as they move away from the emitter in most

circumstances, the geometry of the objects brings about different degrees of shading

and illumination which are highly stochastic. This means linear regression of the spatial

distribution of the location fingerprints is almost impossible.

Due to the fact that the simulation results heavily depend on the environmental

settings which usually vary in a range of uncertainty, a location fingerprint has to

be tolerant to errors. In order to reliably map the location to the fingerprint, the

60

fingerprints must be distinct for every location even in the presence of errors.

Moreover, a preferable mapping function should be fast enough to provide real-time

localization. In order to deal with the discontinuity of high-dimensional fingerprints,

Artificial Neural Network is used to generate accurate, fast and dynamic mapping

functions.

Figure 4-6: A schematic diagram of a one-hidden-layer Artificial Neural Network.

Reviewing the literature, we have concluded that feed-forward back-propagation

Neural Network is the most appropriate in dealing with discontinuous data entries

[18, 99]. As shown in Figure 4-6, it typically consists of three layers: the input layer;

the hidden layer (one or above); the output layer. Data to be trained are fed into the

input layer. Each neuron in the hidden layer stores an unlabelled characteristic. The

output layer directly represents the expected training target. All neurons between any

two layers are fully connected. During training, for each connection a series of variables

are constantly being optimized according to the feedback from the output difference.

4.2.2 Training

To demonstrate the effectiveness of using the Neural Network to generate a map-

ping function, we have investigated a typical residential area in the Hoxton district

of London. The area of interest is 700m×550m. Since it is relatively large, we com-

promise the resolution slightly to get a reasonable computation time by assigning the

pixel size as 5m×5m. In total, there are 15400 receiver pixels. Some 120 objects with

1376 walls and 123 trees are downloaded from the OpenStreetMap. Without precise

knowledge on the sources in this region, we assumed six radio access points of different

61

narrow band frequencies as listed in Table 4.1.

Table 4.1: Virtual emitters allocated to the Hoxton scenario

Source Frequency(GHz) Power(dBm) Location(m) Height(m)

FM Radio broadcast 0.088 30 (490, 150) 30

DAB radio signal 0.225 40 (110, 170) 20

GSM down link 0.9 10 (50, 500) 10

GSM upper link 1.8 10 (550, 480) 10

Wi-Fi 2.45GHz 2.45 20 (315, 340) 2.5

Wi-Fi 6GHz 6 20 (330, 350) 1.5

These sources of different powers are dispersed across the map with the aim of

producing diversity in the fingerprint data distribution. The receiver is always assumed

to be 1.5m above the ground. A sketch of the geometry and Wi-Fi location is shown

in Figure 4-7. Due to the absence of building heights, we assume every roof-top is 15

m above ground.

Figure 4-7: Hoxton district layout with a virtual Wi-Fi transmitter marked as a reddot at (315m, 340m).

We have subsequently simulated the receiver observations using the ray launching

62

model and built a location fingerprint database. The next step is training this database

using the Neural Network. The target outputs of the training are the 15400 pairs of

(x, y) coordinates. The remaining question is how we configure the inputs and Neural

Network parameters.

The RSS plays an important part in categorizing the scenario into hierarchies of

illumination. A 2D view of the RSSs at 2.45GHz is shown in Figure 4-8. Due to

diffraction and wall penetration losses there is significantly less signal strength detected

in shadowed areas than LOS areas. In the LOS areas, the RSS keeps fading farther away

from the transmitter as the free-space propagation loss increases. As can be seen in

Figure 4-8, the multipath and shadowing effect results in remarkable discontinuity of the

RSS distribution which helps to differentiate zones separated by walls. Nevertheless, the

outwardly attenuating envelope of RSS still reveals how far away the receiver location

is from the source.

Figure 4-8: Simulated RSS heat map for the Hoxton scenario provided by a 20dBmemitter of 2.45GHz at (315m, 340m).

The MED and RMSDS are combined to depict the time delay of arrival as well as the

arrival time clustering along the time axis. These properties reflect subtle differences in

which arrivals make up the PDP. They are also expected to help discriminate receiver

locations.

63

Figure 4-9: The most significant AOA x-axis Cartesian projection on the Hoxton sce-nario, where null pixels are assigned -2; transmitter location is assigned 2 marked as ared dot.

In LOS propagation, the AOA provides straightforward information on which direc-

tion the receiver is to the source, as can be seen in Figures 4-9 and 4-10. In addition to

the arrival angle of the strongest signal, we are also interested in exploiting the 2nd and

3rd most significant arrival angles. Since the reception detection guarantees one ray re-

ceived at a time, rays other than most significant arrival provide information on higher

orders of illumination situations. For instance, two different receiver locations might

have identical LOS arriving angles, whereas, their 2nd or 3rd strongest arrivals com-

ing from diverse propagation paths offer evidence to distinguish them. Consequently,

using the top three significant AOAs could not only divide the illuminated region into

‘pizza-slice’ like zones to localize, but help discriminate receiver locations on a smaller

scale.

As a result, the number of columns for a location fingerprint increases to nine as

each AOA contains two scalars. Hence, the row vector of the fingerprint matrix can be

written as

[RSSi,MEDi,RMSDSi,AOA1Xi,AOA1Yi,AOA2Xi,AOA2Yi,AOA3Xi,AOA3Yi]

64

Given M = 6 opportunistic sources, a location fingerprint contains 6×9 = 54 elements.

Hence, the input layer to the ANN is comprised of 54 entry neurons.

Figure 4-10: The most significant AOA y-axis Cartesian projection on the Hoxtonscenario, where null pixels are assigned -2; transmitter location is assigned 2 markedas a red dot.

The Matlab ANN toolbox provides a comprehensive training platform. In addition

to the input and output layers, a series of variables need to be specified. In a feed-

forward back-propagation network the signals flow in one direction from input, via the

hidden layers, to target. Whereas, the error which is obtained from subtracting target

from the current training output constantly back-propagates to the input to reduce the

error for the next iteration. Network type of feed-forward back-propagation is often

cited to produce the fastest convergence and is outstanding in pattern recognition.

The ‘TRAINLM’, short for Levenberg-Marquardt training algorithm, is acknowl-

edged as one of the most accurate and fastest converging functions [115]. It usu-

ally outperforms other training functions in both curve-fitting and pattern recogni-

tion problems. The downside of ‘TRAINLM’ is a larger memory space requirement

which also results in longer execution time. Other popular training functions include

‘TRAINRP’ (Resilient backpropagation algorithm), ‘TRAINSCG’ (Scaled Conjugate

Gradient backpropagation algorithm), ‘TRAINBFG’ (BFGS quasi-Newton backpropa-

65

gation algorithm). We have examined their performance which either exhibited limited

accuracy or lacked stability in achieving a good result in comparison to ‘TRAINLM’.

The adaption learning function, ‘LEARNGDM’ (Gradient Descent with Momentum

weight and bias learning function), is recommended for the incremental training [1].

‘MSE’, short for Mean Square Error, is used to measure the performance. Detailed

network configuration properties are captured in Figures 4-11 and 4-12.

Figure 4-11: Neural Network user interface panel.

Figure 4-12: Neural Network settings.

According to the literature on ANN [124], multilayer Neural Networks are deemed

more capable of handling discontinuous entries. As a compromise between speed and

accuracy, two hidden layers and one output layout are assigned. Demuth [26] suggest

that over half the neurons of the last layer should be allocated to a hidden layer to

achieve better convergence performance. Following the approach of [26], in the Hoxton

example, we assigned 48 neurons to the first layer and 32 to the second. The sum of

the weighted inputs and the bias from the last layer forms the input to the transfer

66

function which generates the output for each neuron. Note that the selection of transfer

function for each layer effects how the weights and biases proceed to reduce the error

in every iteration. Table 4.2 lists the characteristics of common transfer functions.

Table 4.2: Characteristics of transfer functions in the Neural Network

Name Expression Output range Recommendation for use

Log-sigmoid a = logsig(n) (0, 1) Multilayer networks

Tan-sigmoid a = tansig(n) (-1, 1) Pattern recognition problems

Linear a = purelin(n) (−∞, ∞) Function fitting problems

Based on trial and error, we found that using Log-sigmoid for the first layer and

Tan-sigmoid for the second yielded the best convergence performance. A schematic

diagram of the configured ANN for the Hoxton example is shown in Figure 4-13.

Figure 4-13: Neural Network configuration diagram.

Using this configured ANN, we are able to train the input fingerprints for 15400

iterations, equivalent to the total number of receiver candidates, with respect to the

2D coordinates of the required locations. The training result panel is displayed in

4-14. It is worth noticing that Mu represents the training gain of the outcome error

compared to the initial error. This particular training process took six iterations and

was performed within two minutes. The mean square error converged to 169m2 which

implies a mean mapping error of 13m. Considering the fact that a single receiver pixel

takes up a 5m×5m square, the mean error is good.

It is definitely true that by setting a more critical training gain before stopping one

could possibly obtain an even better training performance. However, an overly trained

system may produce worse estimation in practice because the intermediate inputs may

find erroneous outputs even though the 15400 samples are perfectly fitted. A closer

look at the accuracy performance is given in the next section.

67

Figure 4-14: Training result panel for the Hoxton example.

4.3 Accuracy and reliability

Certainly, a theoretically minimal error is not the only goal of a localization algo-

rithm. A variety of factors which might affect the training performance need to be

studied, and it is questioned if this method is applicable to any urban scenario. More-

over, the mapping function has to be resilient to noise and deviations applied to the

input since observed fingerprints are likely to differ from the simulation results which

are used as training data. To find explanations for the above concerns, we carried out

experiments by manipulating variables for the Hoxton district example and have drawn

conclusions based on our observations.

A very straightforward means to improve the accuracy is to make the performance

measure, Mean Square Error, as small as possible. This involves several manageable

factors, such as the selection of training inputs among RSS, TOA and orders of AOA,

the number of opportunistic sources, the resolution of allocated receiver pixels, the

number of layers and neurons in each layer, the transfer functions, the expected training

gain and so on.

68

4.3.1 Number of opportunistic sources

In the previous Hoxton example, we chose six opportunistic sources, specified in Ta-

ble 4.1, as location fingerprint generators and gained 13m of mean error after training.

Figure 4-15 shows the relationship between the mean distance errors and the number of

sources available. Generally, as more opportunistic sources are incorporated the mean

error decreases, whereas, the accuracy gain starts to drop when the number of sources is

greater than six. Up until ten opportunistic sources, enriching the location fingerprint

by incorporating more makes little contribution to the overall mapping performance.

Therefore, it is a matter of balance between the number of sources and complexity of

a location fingerprint to obtain a reasonable training performance.

Figure 4-15: Mean error distance as a function of number of opportunistic sourcesavailable.

4.3.2 Location fingerprint components

We defined a location fingerprint should comprise RSS, MED, RMSDS and top

three AOAs because their spatial distributions may be used to discriminate locations,

69

and that they are measurable. Nevertheless, the significance of each primitive requires

further investigation. To measure the contributions of these primitives, we excluded

one component at a time and trained the rest. Table 4.3 lists the exclusion conditions

of fingerprint primitives and their corresponding mean errors. It can be seen that the

exclusion of the most significant AOA has the most severe impact on performance,

followed by the RSS. Time delays and higher orders of AOA are less crucial but still

result in remarkable deterioration compared to the original 13m error.

Table 4.3: Mean errors against combinations of location fingerprint components (Y isincluded; N is not included)

RSS MED RMSDS AOA1 AOA2 AOA3 Mean error(m)

N Y Y Y Y Y 24.0416

Y N N Y Y Y 22.0000

Y Y Y N Y Y 28.9310

Y Y Y Y N N 22.5832

4.3.3 Location target resolution

During the simulation, we assigned a grid of 5m×5m pixels across the scenario and

trained a location fingerprint for each pixel. Thus, the expected resolution can reach

up to 5m. In this case, only one sample is being trained to identify one target. It

is worth studying whether providing more samples for one target but sacrificing the

resolution would enhance the training performance.

In Figure 4-16, mean errors are plotted against the resolution which is equivalent to

the length of the side of each target square. For example, a 10m×10m pixel contains

four receiver grid points such that their centroid will be used as the target in which

four sets of location fingerprints are trained towards. A 50m resolution suggests that

100 sets of training input are available for one location target. It can be seen that the

performance degrades due to the spatial averaging and blurring of the fingerprint data.

From the trend of mean errors, we conclude that sacrificing the resolution to allow

more training resources does not benefit the performance at all. Moreover, as the reso-

lution increases the mean error also keeps increasing in general. It is not complicated to

interpret that taking the centroid of a group of receivers is indeed wiping out individual

information contained in each receiver location. Hence, the Neural Network is deemed

to be capable of correctly handling the information being trained. It is not necessary

for us to pre-process the input.

70

Figure 4-16: Mean error distance as a function of the pixel side length used in a traininglocation.

4.3.4 ANN configuration

The number of neurons for each layer influences the performance as well. A two-

layer Neural Network has been chosen because it is believed to be more capable of

handling nonlinear entries. Table 4.4 shows the network convergence performance,

measured in metres of mean error, as a function of number of neurons in the 1st layer

and 2nd layer. The range of varying neurons in each layer is set around half of its

previous layer according to [26]. The result indicates the combination of 28 and 20

performs the best.

71

Table 4.4: Mean Error (m) against number of neurons in the 1st layer and 2nd layer

XXXXXXXXXXXXXX2nd layer

1st layer24 26 28 30 32 34

10 20.668 19.966 19.740 20.390 19.854 18.476

12 19.645 20.545 18.785 19.286 18.950 18.407

14 20.265 18.766 20.027 18.857 20.081 18.720

16 20.122 18.941 17.907 19.741 20.840 19.450

18 20.414 19.412 20.458 18.663 20.646 18.383

20 18.658 19.684 17.736 18.156 18.396 18.806

Specifically, as the number of neurons in the 2nd layer increases the mean error

mostly keeps declining and the 1st layer choice seems to make no difference. Generally,

we realize that the more neurons allocated the better accuracy there is, though more

neurons always takes an excessive amount of iterations and processing time. Through

exhaustive searching, we determined that allocating 48 neurons in the 1st layer and 32

in the 2nd gives the smallest error, 13m. Since ANN training process is not supervised,

it may produce networks of different mean errors given the same settings. Therefore,

13m of mean error is merely a nice figure which does not represent an optimised result.

Literally, the pure linear transfer function generates linear response which was found

not to be suitable. The characteristics of common transfer functions are summarized

in Table 4.2. Through trial and error, we determined that using logistic sigmoid for the

first layer and tangent sigmoid for the second provided the best training performance.

Table 4.5: Training performance as a function of the assignment of training gain, Mu

Mu Number of iterations Time(s) Mean error(m)

0.01 6 99 13.0384

0.001 6 88 13.0000

0.0001 6 93 13.2665

0.00001 6 95 13.4536

We also tuned the training gain to reveal its effect on the performance. Table 4.5

shows how the training results are effected by the value of Mu. Since the number of

iterations, processing time as well as the mean error are roughly the same, a conclusion

is drawn that the training gain is not the saturating factor that ceases further training

of the Neural Network. In other words, the Neural Network configuration is no longer

able to provide better mapping performance.

72

4.3.5 Noise level during measurement

In order to demonstrate the robustness of the mapping function, we deliberately

added different levels of AWGN (Additive White Gaussian Noise) to the simulation

results, as well as altered the environmental settings to generate corrupted fingerprints.

Using these corrupted fingerprints, we attempted to prove that the user observations

in the field do not have to be perfect to determine satisfactory localization.

AWGN, i.e., normally distributed random noise, is intentionally added to all 15400×6

RSSs of the original location fingerprints in the Hoxton example. In the first trial, the

mean of AWGN was set as zero to generate noise without bias. The standard Deviation

(SD) of the noise was tuned from 0.01 to 0.03 which equals 2.4dB AWGN (the trained

RSS values are truncated between -140dBm and -60dBm and normalised from -1 to

1) to reveal the impact of noise strength levels. As can be seen from Figure 4-17, the

matching accuracies under SD = 0.01 of noise are marked as a solid blue line in which

approximately 90% of localizations are within 15m of the true locations. At exactly

15m of tolerance, the accuracy performance degrades from 75% to below 60% as noise

strength increases from 0.8dB to 2.4dB on the solid yellow line and dashed magenta

line.

Figure 4-17: Cumulative Distribution Function (CDF) of localization accuracy againstthe standard deviations of the AWGN applied to the fingerprints.

73

Figure 4-18: CDF of localization accuracy against measurement biases added to thefingerprints.

Matching accuracy under 1dB of biases are plotted in a solid blue line as shown

in Figure 4-18. When different levels of biases up to 5dB are used to corrupt the

fingerprints, the mean error performance is degraded from 15m to 20m, and up to

nearly 30m at 90% acceptance rate. Generally, the mapping function proves to be

immune to a degree of noise and biases added to the RSS observations.

4.3.6 Environmental factors

Errors and deviations may be attributed to a limited representation of the environ-

ment. For example, the building shapes and surface properties are usually not available

in full detail. The presence of traffic and pedestrians can also influence the channel.

Furthermore, a reliable localization algorithm has to demonstrate universality in any

urban environment.

In pursuit of answers for the above questions, we have applied the fingerprint ex-

traction and training process to three different scenarios as listed in Table 4.6. In

addition to the Hoxton example, the seven-object virtual scenario in Figure 3-12 and

the Arc de Triomphe area in Figure 3-4 have been studied. Pixel size of 1m×1m was

assigned in the former. As in the Hoxton scenario, 5m squares were allocated in the

Arc de Triomphe scenario since the scale is larger. Six opportunistic sources were used

74

in all cases such that the location fingerprint dimension is always 54.

Table 4.6: Training specifications and performances on different scenarios

Scenarios Hoxton district Seven objects Arc de Triomphe

Scale(m) 700× 550 120× 120 1000× 400

Number of pixels 15400 14400 16281

Number of buildings 120 7 149

Number of sources 6 6 6

Mean error after training(m) 13.000 0.603 30.480

Type of disturbing variable A B C

Disturbed mean error(m) 1.791 2.686 38.508

After training, 0.603m mean error is seen in the seven-object scenario; 17.136m

is seen in the Arc de Triomphe scenario. We then added different types of error per-

turbations to the simulator to generate corrupted fingerprints, and fed them into the

pre-trained Neural Networks (error free). The returned estimations were then compared

with the authentic locations.

Figure 4-19: In test A the material of the wall highlighted in blue is manipulated; intest B the folded wall highlighted in green is moved downwards by 5m.

75

Figure 4-20: In test C a bus, of length 10m and width 3m, painted in solid blueis inserted into the Arc de Triomphe scenario and moved upwards for 10 consecutivesteps.

In test A shown in Figure 4-19, we changed one of the longest surfaces from concrete,

with relative dielectric constant 4.0 and conductivity 0.0001S/m, to glass whose dielec-

tric constant equals 7.0 and conductivity equals 0.005S/m. As a result, the estimation

error rises from 0.603m to 1.791m. In test B, a folded wall was moved downwards by

5m. This time the estimation error increases from 0.603m to 2.686m.

In test C on the Arc de Triomphe scenario, we put a double decker bus at the

junction shown in Figure 4-20. The center of the bus was at (500m, 150m). The shape

of the bus was projected as a rectangular, of 10m×3m, at which its height is equivalent

to average roof height in the scenario. Consequently, the MSE increases from 30.480m

to 38.508m.

In order to demonstrate the trained network is generally reliable to moving obstacles

at arbitrary location, we made the bus move on an upwards trajectory for another 10

steps as shown by the arrow in Figure 4-20. The MSE results relying on the originally

trained network are listed in Table 4.7. The mean error hardly deteriorates for more

than 10m. Notice that as the bus moves away from the main street (Champs-Elysees

Avenue) the mean error actually drops. Presumably, we believe reflections occurring

at a crucial location, e.g., the junction of a thoroughfare, would make a more signifi-

cant difference to the propagation paths which definitely influence the authenticity of

fingerprints significantly.

76

Table 4.7: Mapping performance using the trained network in the presence of a movingdouble decker bus

Bus position on y axis(m) Mean error(m) Degradation(m) Percentage

150 38.508 8.028 26.34%

160 41.201 10.721 35.17%

170 38.515 8.035 26.36%

180 40.021 9.541 31.30%

190 40.884 10.404 34.13%

200 35.986 5.506 18.06%

210 34.937 4.457 14.62%

220 34.989 4.509 14.79%

230 34.844 4.364 14.32%

240 34.823 4.343 14.25%

250 34.834 4.354 14.28%

Nevertheless, the mean error increments among these tests are never significant. It

is generally concluded that the localization algorithm does show a considerable degree

of robustness against various environmental uncertainties.

4.4 Summary

In this chapter, a location fingerprint is firstly defined based upon channel mod-

elling results. These were then applied to ANN to train the location fingerprints to

their corresponding 2D coordinates. This has been shown to generate a deterministic

network leading any user observation towards an estimated location. The accuracy and

reliability of this method have been validated against multiple errors and perturbations

for different scenarios.

77

Chapter 5

Source localization

The localization of sources has drawn increased attentions recently, especially in

high-rise environments where the traditional triangulation methods become unreliable

due to multipath propagation. The prospective applications of such localization tech-

niques range from commercial navigation, criminal investigation, to military defence

systems.

In Chapter 4 we discussed and validated a novel fingerprint-based localization ap-

proach. A ray launching model has been developed and shown to be capable of pro-

viding channel estimations with respect to the observing location. In this chapter, we

will now investigate whether fingerprints of unknown sources can be defined to aid a

source localization solution. An appropriate matching algorithm is required in order

to find the best match within a pre-computed fingerprint database.

This chapter details the idea of a source location fingerprint as well as the develop-

ment of a matching algorithm. Case studies based on real-world environments present

validation of this method, and simulations exposed to a variety of perturbations rein-

force the reliability of this method. Of particular interest in this chapter is the use of

UAVs (Unmanned Aerial Vehicle) for the collection of raw data to be matched. We

also put forward several concepts, particularly the UAV route strategy, to improve the

reliability to meet the needs of practical applications.

5.1 Source localization algorithm

Location of unknown radio sources in urban environments is of great research in-

terest currently. The traditional localization approaches are found to be less effective

in severe multipath environments [31, 41]. Many researchers in the literature have

proposed to determine geographical location of radio sources by monitoring several

78

spatially dispersed sensors [79, 92]. Lohrasbipeydeh et al. [78] proposed a location

decision algorithm using RSS as indicator. Adelipour et al. [5] proposed a source local-

ization method using TDOA and FDOA (Frequency Difference of Arrival). Shahidian

et al. [112] utilized a mobile aerial sensing device to collect spatially dispersed data.

Using a dense network of sensors the localization is essentially transformed into a com-

pressed sensing problem from which a deterministic solution can be found. However,

instalment of the sensing devices becomes a pre-requisite, and the quality of the data

becomes critical which is usually non-ideal in busy city streets [128].

To mitigate the above issues, here we propose to make use of a mobile payload

carried by a UAV, to conduct continuous measurements along ‘street canyons’. From

these measurements, characteristic fingerprints can be extracted to discriminate po-

tential source locations. The key steps to accomplish this source localization are the

development of a practical measuring scheme and a robust matching algorithm.

5.1.1 Data collector

In order to capture the spatial electromagnetic features that discriminate source

locations, the joint contributions of as many pieces of data as possible from the field

measurement is preferred rather than relying on persistent observations in fixed loca-

tions. An example of such data collection strategy is illustrated in B of Figure 5-1.

This is also a more realistic assumption for the ray launching model accuracy since

trends of variation are easier to predict than an absolute channel response.

Figure 5-1: A shows a conventional compressed sensing method in which sensors,marked as purple triangles, are dispersed across the scenario at fixed locations; Bdemonstrates a data collection strategy along a specified path which contains morespatial dispersion information of the received parameters.

Following the idea of remote sensing [122], we note it is indispensable to have a

79

mobile measurement device which is able to go along a specified path with constant

speed and wide coverage. A small drone is, therefore, an efficient tool to collect much

information in a short period [37, 126]. By careful selection of the flight path which may

be constrained due to building height or for security reasons, a data series comprised

of electromagnetic observations can be recorded to leverage the fingerprint extraction.

Figure 5-2: An illustration on how to use a UAV to collect data from the field: theUAV moves along a route marked in purple in the urban environment; its observationscan be simulated using the ray launching model given transmitter source candidates(Tx1, Tx2, Tx3, ... Tx9).

The fingerprint obtained above would be an input enquiry to the algorithm which

seeks to determine the source location during the “n-line” query phase. The “off-line”

job is one of exhaustively simulating what the drone would see along that path with

respect to a grid of source locations. From these simulations, a look-up table can be

constructed forming the fingerprint database for all candidate locations.

As can be seen from Figure 5-2, a quad-rotor drone can be used to carry out a

measurement campaign along a route designated by the purple arrow. The blue network

of grid locations implies that the electromagnetic characteristics at each location can

be simulated from the ray launching model. The red dots in the upper layer denote

the emitter candidates which are very likely to generate disparate observations along

the route. In principle, it is the simulated observations along that route which lead to

fingerprint extraction, so as to give rise to source localization.

Hence, the UAV is playing the role of data collector by conducting continuous

sampling ideally at a constant speed and at a fixed height. A desirable route is expected

to wind through street canyons and provide a wide-spread coverage in which the UAV

80

can ideally thread across LOS and NLOS areas from different perspectives to create

variations in the received parameters. Further discussion on route selection is given in

section 5.2.3.

5.1.2 Feature transformation

Given a time series of simulated receiver observations along a specified UAV route, a

link between the received characteristics and transmitter source locations is made. Yet

this alone is not adequate to make a source localization algorithm unless the features can

be explicitly and numerically recognized. A feature transformation process is necessary

to quantize both the UAV observation and the simulated observations [53].

As has been discussed in Chapter 4, a location fingerprint may contain a combi-

nation of normalized RSS, TOA and AOA which are derived from Eqn. (4.1), Eqn.

(4.2), Eqn. (4.3), Eqn. (4.4). A typical field measurement of RSSs is shown in Figure

3-16. Superficially, variations over short distances are likely to be attributed to noise

and fast fading which are less valuable, whereas, variations over longer distances tend

to reveal genuine information on path loss.

A significant complicating factor is that the UAV may not be perfectly synchronized

in time and position. That is, the route may deviate from the expected one. For

example, suppose the UAV measures some parameter with a sinusoidal variation such

that time variation is shown in A of Figure 5-3. Consider now, a time delay during

measurement results in a phase shifted red curve in B. Inaccurate attenuation or surface

reflectivity assignment may make amplitude distortion apparent as shown in C, while

noise and fast fading would corrupt the curve as shown in D.

To counter these problems, the basic idea of feature transformation is to project

the observed data series onto a lossless domain in which periodic characteristics are

emphasized, whereas, the phase and amplitude can be stripped.

A Discrete Fourier Transform (DFT) happens to be such a lossless transform and

is able to decompose a data series into a number of periodic components. The k -th

DFT coefficients, Xk, can be written as;

Xk =N−1∑n=0

xn · e−j2πkn/N (5.1)

where N is the number of samples and xn represents the n-th sample value. Both real

and imaginary parts of the frequency components are taken as feature eigenvector.

81

Figure 5-3: An illustration of potential data time-series corruptions: case A shows theoriginal waveform in blue; case B shows the waveform after phase shift which is markedin red; case C shows the waveform after amplitude distortion which is marked in red;case D shows the waveform exposed to AWGN which is marked in red. (Scales are inarbitrary units)

After applying the DFT to reveal spatial frequency features, a Low-Pass Filter

(LPF) can be used to reduce or eliminate high frequency components which are most

possibly caused by noise and fast fading. A simple example of such LPF may have

a gain equal to unity below the cutting-off frequency and zero above. The filtered

spectral components form the feature vectors for the specified route.

Through these steps, the features of the source locations are transformed into lower

frequency coefficients in frequency domain from the original time-series. These feature

vectors are stored in a look-up table to be checked against an input vector which is to

be obtained from field measurement through an identical measurement procedure.

5.1.3 Source location fingerprint

In order to identify the source location accurately, the ‘fingerprint’ database, which

contains feature eigenvectors extracted from the simulated UAV observation, should

be built for a sufficiently large number of source candidate locations. These source

candidates are also allocated in a grid-like fashion with a uniform separation distance

equivalent to the localization resolution [20]. For each simulated source location, the

82

discrete observations along the UAV route should be concatenated into a data series.

We aim to demonstrate that these electromagnetic observations are not only measurable

but possess physical meanings that can be used to discriminate transmitter locations.

Figure 5-4: Simulated UAV route for the Hoxton scenario where the route is highlightedin yellow and buildings are depicted in dark blue.

With reference to the Hoxton district, an example UAV path is highlighted in

yellow as shown in Figure 5-4 (right). The high-rise buildings are depicted in dark blue

over which it is forbidden to fly the UAV. The green pixels show many obvious street

canyons. Obtaining all receiver characteristics allows more flexible route planning as

well as freedom in analysing an optimal route in terms of the feature transformation

process.

By performing the simulation and extraction process for all source locations, a

source location fingerprint can be determined eventually consisting of the lower spatial-

frequency components which contain the varying trends of RSS, MED, RMSDS and

AOAs. An input measured fingerprint is obtained following the same approach and used

to search against the simulated fingerprint database to estimate the source location.

The final step to success is a matching algorithm to find the best match between the

measured fingerprint and those simulated fingerprints in the database.

5.1.4 Matching algorithm

A commonly used algorithm that could be applied to measure eigenvector similarity

is the Euclidean distance. It considers two n-component vectors as two points in n-

dimensional space and calculates the root mean square of their distance in that abstract

space [58]. In a Cartesian coordinate system, an expression of the Euclidean distance

83

between −→a = (a1, a2, ..., an) and−→b = (b1, b2, ..., bn) can be written as;

‖−→a −−→b ‖ =

√√√√ n∑i=1

(ai − bi)2 (5.2)

where n denotes the dimension. Other common distance measures include discrete

metric, Manhattan distance metric (also referred to as Taxicab distance), graph metric

etc. These metrics have been examined by using a measurement-based fingerprint to

find best match among the database of simulated ones.

However, disappointingly, the test results show that a simple Euclidean distance

metric, or the above metrics, do not yield an accurate matching algorithm. Both

measurement errors and flaws in these approaches may account for such failures. Mea-

surement errors include the interferences of noise and fast fading, measurement route

deviation due to wind effects, poor synchronization as well as the fact that the sampling

may not be evenly distributed along the route as expected.

A critical mechanism flaw with the above metrics concerns the significance of each

dimension. Whereas, in reality the unexpected and unknown phase and amplitude

distortions permit the bins to shift along the frequency axis, which makes dimensional

alignment problematic. To mitigate this problem, the ideal matching algorithm must

possess a flexible pattern recognition scheme which is tolerant of distortions and mis-

alignments.

According to [56, 72], the Dynamic Time Warping (DTW) approach is ideally

suited to this kind of problem. Given two data series −→a = (a1, a2, ..., an) and−→b =

(b1, b2, ..., bn) of length n and m, respectively, an n-by-m matrix can be constructed

in which the (i, j) element contains the distance between two points ai and bj . Each

matrix element corresponds to the alignment between points ai and bj . A warping path

W is a contiguous set of matrix elements that define a mapping between −→a and−→b .

The k -th element of W is defined as wk = (i, j)k. Hence, the DTW is the path that

minimizes the warping cost given by

DTW(−→a ,−→b ) = min

{∑k=1

wk

}

If the fingerprints to be compared are regarded as data series −→a and−→b , the path

that leads to the minimum warping cost corresponds to the best alignment. Then

Euclidean distance, or other distance metrics, can be applied to calculate correspond-

ing distance. It should also be noted that DTW can deal with inputs of different

lengths. Further note that an important variable of DTW, named as width of adjust-

84

ment window, restricts the warping path to be within a specific number of samples of

a straight-line fit. The effect of distance metric used and width of adjustment window

in terms of DTW are discussed in sections 5.2.1 and 5.2.2, respectively.

Figure 5-5: A comparison between Dynamic Time Warping and Euclidean matching.

Specifically in our fingerprint recognition problem, the DTW algorithm is able to

exhaustively seek for the best match from a discrete data among its neighbours, so as

to determine an optimal alignment between the two data series. An illustration on how

DTW works superior than Euclidean matching is shown in Figure 5-5. Using Euclidean

distance, the blue envelope is going to find a large difference to the red. On the other

hand, DTW overcomes the stretching distortion and acknowledges their similarity.

In this final step, the DTW is assisting us in retrieving similarity measures between

the sought fingerprint and simulated source location fingerprints. The best match will

be returned as the estimated source location. Further accuracy and reliability analysis

are detailed in the next section.

85

5.2 Accuracy of analysis

A data mining approach for fingerprint extraction and localization has been pro-

posed. Nevertheless, the practical situations are rarely as straightforward. Several

factors may undermine the effectiveness of the approach, e.g., limited accuracy of the

ray launching model even though it has been validated, uneven sampling positions along

the route, wind conditions during measurement, parameters of the matching algorithm,

as well as the route selection strategy which generates the fingerprint. These factors can

be generally divided into two categories: the noise and deviations during measurement

and simulation; the UAV route selection which triggers location determination.

Table 5.1: Case studies for two real-world scenarios

Case study Hoxton district Munich downtown

Scale 750m× 550m 2400m× 3400m

Number of receiver pixels 15400 54471

Pixel size 5m× 5m 10m× 10m

Number of objects 120 2088

Number of walls 1376 34890

Number of trees 123 1758

Number of candidate emitters 130 616

Number of samples 234 970;355;1031

Expected resolution 50m 100m

To be studied Noise & perturbations Route strategy

In order to verify the source localization theory, we carried out tests on two real-

world scenarios. As is shown in Table 5.1, a typical residential area in Hoxton district,

London, and the Munich city center area were investigated. Both of them come with

serried buildings and trees while having broad coverage with precise simulated receiver

pixels. In this study, 13×10 = 130, and 22×28 = 616 emitter candidates, respectively,

were assigned in grids across the scenarios to give resolutions of 50m and 100m. In

the Hoxton district, the aim is to mainly concentrate on the accuracy performance in

the presence of various perturbations. In the Munich scenario, the aim is to primarily

study which one among the three measurement routes is the best for fingerprinting and

how to select an optimal UAV route in any urban environment.

86

5.2.1 Measurement error and path deviation

A UAV route winding through street canyons is marked in yellow as shown in Figure

5-6 (a). The UAV is assumed to have a constant velocity of 1m/s and a sampling rate

of 0.2 samples per second. Thus, it is able to take a sample every five metres which

guarantees a continuous sampling trace comparable to the receiver resolution. Since

there are 130 candidate emitter locations of which each has a signature fingerprint,

we would like to test how much degree of interference a fingerprint could bear to still

successfully determine the true emitter location.

The scenario where the UAV only uses half the sampling rate (0.1 samples per

second) which gives rise to skipping samples is shown in Figure 5-6 (b). In Figure 5-7

(c), the first turning way point is slightly moved towards the right to imitate the wind

effect blowing to the right during first half of the journey. In Figure 5-7 (d), the UAV

just keeps making zig-zag sampling decisions along the path to imitate the presence of

random directional gusts which unexpectedly deviate the measurement route.

Equivalent perturbations were applied to the originally simulated electromagnetic

primitives in order to create corrupted observations which give rise to noisy and de-

viated fingerprints. The accuracy performance is a measure of how many among the

130 corrupted fingerprints can still be used to determine the authentic source loca-

tions. Table 5.2 lists the accuracies against different kinds of perturbations in which

the observed RSS, TOA and AOAs are all subject to 4dB AWGN.

Table 5.2: Accuracies of relying on RSS, AOA1X, AOA1Y, MED or RMSDS subjectto various simulated measurement errors

Accuracy(%) RSS AOA1X AOA1Y MED RMSDS

AWGN=4dB

Directional wind 95.38 62.31 67.69 66.15 63.85

Random wind 92.31 50.00 48.46 53.08 54.62

‘Manhattan’ metric 100 100 100 96.15 94.62

Half sampling rate 100 100 100 94.62 90.00

Half LPF cut-off 100 100 100 96.15 88.46

87

Figure 5-6: Simulated UAV route deviations applied to the Hoxton scenario: a) Originalroute; b) Halved the original sampling rate.

88

Figure 5-7: Simulated UAV route deviations applied to the Hoxton scenario: c) Direc-tional wind blowing where the first turning way point is shifted; d) Random-directionalwind blowing when the UAV is making zig-zag sampling decisions along the route.

89

Note that the accuracies are obtained in Table 5.2 entirely relying on one observed

electromagnetic primitive at a time. It can be seen that the RSS appears robust to

all kinds of corruption compared to the other primitives. The RSS remains over 90%

accuracy in either kind of windy condition, and returns perfect estimations for all types

of error. Except for wind effect, the trends of variations of AOA1X or AOA1Y are also

found to be reliable. Whereas, time delays are not very good fingerprint primitives to

discriminate source locations.

In fact, the behaviour of the accuracy is to be expected because the fingerprint ex-

traction method has been, in a sense, specifically tailored for RSS. Imagine a boundary

of LOS and NLOS areas separated by the shadow of a wall. The RSS variations of

variation tend to be smooth and continuous due to diffractions on the vertical edges of

the wall. The feature transformation utilizes DFT and a LPF to extract large spatial

variations as a fingerprint eigenvector which favours continuously changing primitives

such as RSS. The TOA and AOA variations possess more discontinuities in compari-

son. On the other hand, it is usually the easiest to measure RSS with simple mobile

equipment fitted on a UAV.

It can clearly be seen that windy conditions have the most dramatic impact on

accuracy performance compared to other variables. Under randomly blowing wind, only

the RSS stays reliable. All primitives other than RSS are severely affected by directional

wind, they still provide good information on the true source locations. Using a hybrid

method, e.g., KNN, combining multiple parameters including AOA1X, AOA1Y, MED

and RMSDS may be a promising approach to better determine the authentic source

location.

In section 5.1.4, we have introduced a DTW algorithm in which a Euclidean dis-

tance metric is used to calculate similarity between two realigned data series. The

‘Manhattan’ metric has also been studied. The results show that the localization ac-

curacy remains outstanding. Additional distance metrics have been studied including

the modification of the power of (ai− bi) in Equation 5.2 from 1 to 3 in 0.1 increment.

These modifications have little effect on the results. Note that the fingerprint elements

defined can be non-physical such that the matching procedure does not have to make

physical sense but can only be judged good or bad by its resulting accuracy.

Halving sampling rate is, in essence, getting rid of every other discrete measurement

sample. In the frequency domain, exactly half of the bins are available this time and

the higher frequency components are omitted. Therefore, it is actually a question of to

what distance does a spatial variation make a difference.

Applying a LPF with one half of the cut-off frequency gives a similar result. An

ideal LPF will only keep the DFT coefficients below the cut-off frequency. Given the

90

sampling frequency here Fs =0.2Hz and the total number of samples along the route

N = 234, each frequency bin can be written as

fbin =FsN

= 8.55× 10−4Hz

If n is the cut-off frequency which equals the number of bins included and the UAV

speed is v =1m/s, the smallest detectable distance, D, of variation becomes

Dmin =v

n · fbin≈ 1170

n

Originally only spatial wavelengths of over 50m are taken as a fingerprint signature,

so solving for n gives1170

n≥ 50

n ≤ 23.4

Thus, 23·fbin is set as the LPF cut-off frequency. If it is halved, the 11 lowest frequency

components are taken as a fingerprint. As can be seen in Table 5.2, the results show

that halving the LPF cut-off frequency does not significantly affect the performance.

Figure 5-8: Light blue envelope represents measured fingerprint of route ‘METRO 200’;blue one is the corresponding best match found among the fingerprint database. Ascan be seen significant amplitudes are all located at the lower frequency end.

The effect of varying the cut-off frequency has been investigated by considering as

few as two to up to 100 frequency bins. By way of example, Figure 5-8 shows simulated

91

and measured spatial frequency amplitudes along a UAV route (further details of the

route and measurement are given in section 5.2.3). It can clearly be seen that only the

lowest frequency components have significant amplitude. Typically it has been found

that as long as more than five bins are included the accuracy is always very good.

This implies that spatial variations over 11705 =234m make the greatest contribution to

source location discrimination. Furthermore, the absolute amplitude values of higher

frequency components are in fact much smaller than lower ones. Even though small

spatial variations down to 1170100 =11.7m wavelength are taken into a fingerprint, the

high frequency fingerprint elements make little difference to the matching result.

5.2.2 Noise level and bias

Deviations may not only come from measurement errors but also the inevitable

background noise and environmental perturbations. Path loss factors, reflectivity of

building surfaces, fast fading, unexpected obstacles etc. cannot always be accurately

represented. The localization algorithm must possess a degree of robustness against

noise and biases contained in the observations.

Figure 5-9: Accuracies of relying on RSS, AOA1X, AOA1Y, MED or RMSDS againstnoise strengths; the curves represent the average of 100 simulations, and the error barrepresents the minimum and maximum.

92

Figure 5-10: Normalized accuracies of relying on RSS, AOA1X, AOA1Y, MED orRMSDS against biases; the curves represent the average of 100 simulations, and theerror bar represents the minimum and maximum.

Figure 5-11: Mean error distances of relying on RSS, AOA1X, AOA1Y, MED orRMSDS against noise strengths; the curves represent the average of 100 simulations,and the error bar represents the minimum and maximum.

93

Figure 5-12: Mean error distances of relying on RSS, AOA1X, AOA1Y, MED orRMSDS against biases; the curves represent the average of 100 simulations, and theerror bar represents the minimum and maximum.

Therefore, we tested the accuracy performance with respect to increasing levels of

noise strength and bias applied to the simulated measurements as shown in Figures

5-9 and 5-10. To generate representative statistics, 100 sets of normally distributed

noise are generated for each level of strength or bias. The curves show the average

values and the errorbars represent the greatest (minimum and maximum) deviations

found. The accuracies were determined solely relying on only one electromagnetic

primitive at a time. It can be seen that the RSS, AOA1X, AOA1Y all demonstrate

impressive robustness in the presence of noise. Under 10dB AWGN, all of them produce

results with over 95% accuracy. The MED is less reliable, whereas, the RMSDS is not

trustworthy at all.

On the other hand, RSS, AOA1X and AOA1Y also perform extraordinarily well

against biases. All of them return over 90% accuracy at 16dB bias which is enormous.

This performance, in the presence of such a large bias, validates the effectiveness of

DTW which is able to recognize similarities between data series regardless of the ampli-

tude gap. The performance of MED is acceptable, whereas, RMSDS is totally ineffective

at the presence of bias.

Figures 5-11 and 5-12 reflect the performances in terms of mean error distance.

Below 10dB of strength, the AWGN is not able to deteriorate the mean error by more

94

than 20m if RSS, AOA1X or AOA1Y is being used. By contrast, RSS, AOA1X and

AOA1Y can tolerate a bias up to 14dB and achieve the same performance.

It can be seen that AOA1X outperformed AOA1Y both in terms of localization

accuracy and mean error distance. Since their values are Cartesian projections of one

vector, one might expect them to have the same performance. However, this is not the

case since the results are route dependent. This has been investigated by reversing the

coordinate system in the Hoxton scenario in which the length, 700m, used to be larger

than the width, 550m. Based on a symmetrical UAV route and identical transformation

procedures, the analysis was repeated. The accuracy and mean error distance against

the increasing level of noise strength for 100 different sets are shown in Figures 5-13 and

5-14. It can be seen that AOA1Y outperformed AOA1X which implies that superiority

of AOA1X compared to AOA1Y depends on the geometry. Since the area of interest

is a rectangular, a parameter which has more significant variation along the longer

side possesses better spatial dependency, allowing improved fingerprint discrimination.

This explains why the Cartesian projection of AOA along the longer side of the scenario

always produces slightly better performance.

Figure 5-13: Accuracy of relying on RSS, AOA1X, AOA1Y, MED or RMSDS againstbiases after reversing the coordinate system of the Hoxton scenario; the curves rep-resent the average of 100 simulations, and the error bar represents the minimum andmaximum.

95

Figure 5-14: Mean error distances of relying on RSS, AOA1X, AOA1Y, MED orRMSDS against biases after reversing the coordinate system of the Hoxton scenario;the curves represent the average of 100 simulations, and the error bar represents theminimum and maximum.

The width of adjustment window of the DTW specifies the number of neighbour

samples to be considered along the warping path which may affect the fingerprint

matching accuracy. Therefore, an investigation is carried out by increasing the width

of adjustment window from 1 to 6, which is applied to RSSs in the presence of 8dB

AWGN. As can be seen from Table 5.3, the average accuracies are very similar over

100 simulations. The best accuracy, 96.84%, is obtained when the width of adjustment

window equals three. Since when width of adjustment window equals one the accuracy

still remains close to the optimum, it can be concluded that the mismatch between

corrupted fingerprints and the authentic fingerprint is usually no more than one sample.

Table 5.3: Average accuracies relying on RSS over 100 simulations when the width ofadjustment window for the DTW algorithm increases from 1 to 6. (8dB AWGN isapplied to the simulated RSSs)

Width of adjustment window 1 2 3 4 5 6

Accuracy (%) 96.50 96.75 96.84 95.73 96.24 95.81

Generally, it can be concluded that MED or RMSDS measurement alone is not

capable of source localization. Variations of RSS and the most significant AOA collected

by a UAV have the potential of reliably revealing source locations. The width of

96

adjustment window in terms of DTW is less significant for the successful matching of

fingerprints in this localization approach. However, there are yet doubts on whether

any UAV path in the environment is capable of telling source locations, and what kind

of route strategy determines the best localization accuracy. These are considered in

the next section.

5.2.3 Route selection

Figure 5-15: Measurement campaigns in Munich downtown area with three routes,named as ‘METRO 200’, ‘METRO 201’ and ‘METRO 202’.

We have been using the COST 231 action measurements from Mannesmann Mo-

bilfunk GmbH [36] to validate our 2.5D ray launching model in Chapter 3. In this

97

dataset, there are three different measurement routes available with an approximately

constant sampling rate as shown in Figure 5-15. During measurement, the equipment

was held 1.5m above ground and the separation distance between samples is 14m on

average. Square pixels of length 10m have been allocated across the map which are

adequate to simulate the observations.

We could simply assume the UAV is flying 1.5m above ground while getting path

loss measurements as shown in Figure 5-16. Thus, the source localization algorithm

can be applied to these three routes to test their performance.

As stated in Table 5.1, there are 970, 355, 1031 samples along routes ‘METRO 200’,

‘METRO 201’ and ‘METRO202’, respectively. We were able to simulate 28×22 = 616

emitter candidates in a same configuration as the real one which transmits a 1W

narrow-band signal at 947MHz through omni-directional antennas. These emitters are

assigned in a fashion of grids with 100m uniform separation.

Figure 5-16: Pathloss trends along the three measurement routes on the Munich sce-nario.

Using the ray launching model, we obtained the RSS distributions with respect to

all emitter candidate locations. Figure 5-17 shows the estimated RSS heat map given

the source location at (1281.36m, 1381.27m). The trajectories of the measurement

campaign are plotted against the simulated heat map as shown in Figure 5-18.

98

Figure 5-17: Simulated RSS heat map given a 947MHz radio source transmitting 1Wpower at (1281.36m, 1381.27m) on the Munich scenario.

The closest grid point at (1300m, 1400m) is expected to be the best match. Com-

parisons between simulated RSS trends when the emitter is at (1300m, 1400m) and

the measured ones can be found in Figures 5-19, 5-20, 5-21.

99

Figure 5-18: Measurement routes against RSS heat map given a 947MHz radio sourcetransmitting 1W power at (1281.36m, 1381.27m) on the Munich scenario.

As has been shown in section 3.5, the agreement between measurements and simu-

lations is good (Mean Absolute Error is between 5.7dB and 7.7dB) and the correlation

coefficient between simulations and measurements is high (over 0.8725 for any mea-

surement route). Importantly here the deep fades are well represented which gives

optimism that the spatial frequencies can be reliably predicted.

100



101


Following the source localization procedures, we established fingerprint databases

based on simulated RSSs along the three different routes. Each series of measurements

can be regarded as an entry to derive a UAV observed fingerprint. This fingerprint

is subsequently exposed to increasing levels of noise strength and bias, and searched

against the database using the proposed matching algorithm.

Figure 5-22 exhibits the localization error performance against noise strength and

bias where darker colours represent a smaller error and brighter colours suggest a larger

error. The dark blue indicates ‘spot on’ in which the closet source location is found.

Light blue indicates an error of less than 150m, i.e., a neighbour of grid (1300m, 1400

m) is located. The yellow and red dots are worse estimations, being at least 200m

away.

102

Figure 5-22: Localization errors for the three routes against AWGN strength (standarddeviation) and bias (mean) where brighter colours indicate large errors and darkercolours indicate small errors.

103

There are several interesting characteristics seen in the error distributions. Firstly,

as long as the AWGN strength added to the RSS observation is less than 5dB a best

match is mostly returned along any of the three routes. No matter how much bias

there is, the minimum error always holds. This discovery confirms that the DTW is

working very well in realizing similar shapes between two fingerprints while neglecting

the absolute value difference. Since noise and bias added to the fingerprint may shift

amplitudes and phases of the frequency components, the envelope of these low frequency

components, i.e., the fingerprint, sees distortion which, nevertheless, can be mitigated

by DTW algorithm.

The two worse routes, ‘METRO 201’ and ‘METRO 202’, remain relatively robust

to a 5dB noise strength and any large bias. This suggests the source localization

algorithm may be generalized even though the UAV path cannot be freely chosen.

It is clear that the ‘METRO201’ is the worst route among the three. As the noise

strength increases, absurd location estimations are witnessed. The cause of these errors

is mainly an insufficient number of samples. If the sample separations are identical, a

lack of samples indicates a limited travel distance. Thus, it is problematic to extract

spatial frequency components for short distances. In comparison to ‘METRO 201, the

other two routes have double of the number of samples. As a result, ‘METRO 200’ and

‘METRO 201’ can generally make an estimation no worse than 200m in error given an

extremely noisy and biased channel. Hence, these results indicate that a longer UAV

route which provides sufficient spatial frequency components is preferable.

Figure 5-23: Light blue envelope represents measured fingerprint of route ‘METRO200’; blue one is the corresponding best match found among the fingerprint database.

104

Figure 5-24: Red envelope represents measured fingerprint of route ‘METRO 201’; blueone is the corresponding best match found among the fingerprint database.

Figure 5-25: Magenta envelope represents measured fingerprint of route ‘METRO 202’;blue one is the corresponding best match found among the fingerprint database.

Lastly, ‘METRO 200’ seems superior to ‘METRO 202’ although they are of similar

length and both wind through street canyons. We suppose it is the general moving

direction that differentiates them. If a closer look is taken at their fingerprint envelopes

in Figures 5-23, 5-24 and 5-25, it is easy to see that the fingerprint of ‘METRO 200’

in its DFT has one significant low frequency component at the third frequency bin,

105

as well as a series of high strengths at mid-frequency range. Whereas, the signature

of ‘METRO 202’ has considerable ripple in the low frequency region which makes it

vulnerable to misalignment with the target fingerprint using the DTW algorithm.

On the map of Figure 5-15, ‘METRO 200’ which is marked in red and starting from

north part of the map is gradually moving towards south, i.e., the real source position.

This approaching direction is essentially generating a slowly rising trend of RSS which

gives rise to the significant low frequency impulse seen in its DFT. However, ‘METRO

202’ marked in green repeatedly approaches and departs from the source which is likely

to generate a series of less outstanding spatial variations. It is highly likely such ripple

in the low frequency components are mismatched if DTW does not determine a perfect

alignment. Therefore, a UAV route that is definitely approaching or leaving the source

location is preferred. Notice that the source location is unknown in a priori. As a

result, a proper route strategy should be expressed as generally moving towards one

direction without weaving back and forth.

In brief, RSS is the most reliable primitive against various kinds of interferences in

terms of constructing source location fingerprints. A preferable observing route should

generally head towards one direction while having as longer extension as possible.

5.3 Summary

This chapter details the development of the radio source localization scheme. A

source location fingerprint is defined relying on field measurements from a UAV. After

transformation of the measured data, a database is constructed containing fingerprints

of all source candidate locations. A DTW-based matching algorithm is used to find the

nearest source location. This method has been tested exposed to measurement errors,

deviations, AWGNs and different UAV routes. Outstanding accuracy and universality

of use are validated.

Although a longer measurement route is preferable for the case studies in this

chapter, this may not be true if the route extends too long. Since the lower frequency

components which represent spatial variations over a relatively longer distance are more

capable of determining source locations, they would be combined with more measure-

ments to make higher frequency ripples along a much longer route. Whether the spatial

frequency components of a much longer route are still trustworthy to determine source

locations, and how long a route should be to offer the best localization performance

are not clear, and could be part of further work.

106

Chapter 6

Enhancement and program

optimization

The proposed navigation and source localization approach largely relies on the de-

velopment of a propagation model and localization algorithms. Optimization of the

system is a key step to guarantee the success. The fingerprints of locations are inher-

ently static as long as the opportunistic signals and propagation environment remain

unchanged. In Chapter 4, it has been shown that small changes to the environment do

not result in significant errors. Thus, we attempt to share the static fingerprint database

with users such that they are able to navigate locally. In this chapter, a sequence-based

step detection scheme which aims to minimize the communication between user and

server is proposed to assist the navigation system during off-line periods.

For the purpose of building channel fingerprint databases, the 2.5D ray launching

model developed in this work has high requirements for both accuracy and efficiency.

Matlab is an extraordinary platform for scientific research and development. Not only

does it provide easy handling of matrix and structure array, but comprehensive em-

bedded functions on visualization, data manipulation as well as portability to other

software. In some instances, there is a significant performance overhead. The adoption

of Matlab allows for rapid development and high accuracy but not always the highest

speed. In this chapter, the methods adopted to optimize the Matlab master program

are expanded in detail.

The use of GPU based multi-threading computing is deemed an effective means for

acceleration. In this chapter, we briefly explain the parallel computing compatibility

of our program and introduce the advantages as well as limitations of using GPU. The

implementation of this accelerator has been done with the help of CUDA and the speed

up performance has been evaluated. Eventually, above acceleration and optimization

107

methods make accurate and rapid propagation modelling a realistic prospective.

6.1 Sequence-based navigation scheme

One difficulty of fingerprinting techniques that restrict the precision of localization

stems from the reliability of the fingerprint data. Unreliable fingerprints which are

likely to result in poorly estimated locations may come from interference, fast fading

or path obstructions. As the mobile user usually proceeds along a continuous path, a

consideration might be if previous footsteps could be exploited to correct the navigation

solution along the path by avoiding odd location estimations.

This may help with continuity since a mobile device occasionally suffers from lost

signals in densely built areas. Thus, we propose to navigate the user with as little

communication between the mobile device and the server as possible. Given the local-

ization algorithm and preliminary fingerprint database, our proposed navigation system

adopts a built-in sequence detection algorithm adapted from an indoor sequence-based

approach [129, 130].

Based on the assumption a mobile user will go along a continuous path, a package

of neighbour fingerprints is passed to the user among which the next step is expected.

Since 5m or 10m pixel length is usually applied, a 12-neighbour detection (see Figure

6-1) is used to find the next step in a short interval, e.g. 2 seconds. During 2 seconds,

a pedestrian user can be assumed to move for no more than 10m. Consequently,

12-neighbour pixels are promising to identify a position change of up to 15.6m.

Figure 6-1: An illustration of the sequence-based navigation method when positionnumber 2 is the best anticipated next step.

On the mobile device, a simple and reliable algorithm, K-Nearest Neighbours, is

adopted to calculate the closest next step. As is shown in Figure 6-1, if sequence

108

number 2 gives the minimum difference using KNN it is decided as the next step. If

number 0 is the best match, the user is assumed to stay stationary.

The package sent to the user contains all the fingerprint primitives of 12 candidate

locations which may still be too many, and some of them having small weight after

training. Therefore, during the off-line phase principal component analysis is applied to

the database of fingerprints to determine the most significant primitives. For example,

three among the RSS, MED, RMSDS, AOA1X, AOA1Y etc. may be chosen whose

combined weights matter the most. If RSS, AOA1X and AOA1Y are taken, the step

index i is exhaustively searched from 1 to 12 to find the minimum of the expression

below

{w1 · (RSS −RSSi)2 + w2 · (AOA1X −AOA1Xi)

2 + w3 · (AOA1Y −AOA1Yi)2}

where wj represents the assigned weight of j -th selected primitive. To prevent the

accumulation of errors over time, periodically it may be necessary to send a complete

observed fingerprint to the server to correct its current location.

Following this approach, a sequence-based database can be obtained from the orig-

inal fingerprint database for every pixel. Prior to the starting of navigation process,

a package of sequence-based database is delivered to the mobile device, so that the

next few steps of the user can be automatically decided by monitoring the observed

fingerprint against the sequence-based database which is stored locally.

Hence, the built-in navigation algorithm is able to guide the user by tracing an

optimized next step. This method makes most of the localization effort off-line which

effectively reduces the required communication between users and server. This greatly

improves the practicality of the navigation system.

6.2 Matlab

The master program of the 2.5D ray launching model has been developed in Matlab

2016b. The inputs to this program include initiation of the source ray wavefront, layout

of the environment and specifications of the model, such as radius of the reception

sphere, reflection limit, truncation range of RSS etc. The ray launching mechanisms

described in Chapter 3 work together to produce channel estimations. Dependant

variables RSS, TOA, AOA with regard to locations are the outputs which are stored

in a cell array. A flowchart of this master program is shown in Figure 6-2.

109

Figure 6-2: A flowchart of the master program of the 2.5D ray launching model.

A primary unit of the ray launching engine detects the intersection situation be-

tween a vector, i.e., a ray, and a line segment, i.e., a building surface. Depending on

the type of intersection, a reflection processor, diffraction processor or transmission

processor will determine the propagation behaviour onwards. A critical bottleneck

that restricts the performance of CPU-based program is the tremendous number of for

loops, which analyse intersections between one among the equally-radiated rays and

one building surface at a time. In this work, it takes even longer if a fully-connected

mesh of transmitter and receiver relationships is the final objective. The pseudo code

of the intersection detection unit can be written as follows;

110

Algorithm 1: Intersection detection function for determining the ray propa-

gation behaviour

1 Intersection Detection;

Input : Virtual source O(xo, yo); ray −→r ; surface (xa, ya)→ (xb, yb)

Output: [flag, distance]

2 if intersection = true then

3 if intersection{type} = edge then

4 return [flag, distance] = Diffraction( );

5 else

6 return [flag, distance] = Reflection( );

7 go to Transmission(−→r , distance);8 end

9 else

10 return [flag, distance] = set(0);

11 end

For instance, if 720 rays with up to 7 reflections are shot and 34890 walls (which are

from the Munich map occurred in the scenario), to calculate 100 emitter candidates and

54471 receiver pixels it takes about 9.58× 1014 iterations of the intersection detection

unit. A brute force calculation based on a 2.6GHz Intel Core i5 dual-core processor

would take many month to compute.

To reduce the computation time, we have adopted quite a few means of acceleration.

Firstly, we omit calculations which do not make sense in physics, e.g., rays shooting

out of the region, termination of rays diffracted or dropped below a signal strength

threshold. Secondly, we abandon built-in Matlab functions which have a significant

overhead by using simplified math operations instead. For instance, the embedded

cross product function is called many thousands of times. Since we actually only use a

summation of scalar products to obtain what is needed, a lot of time can be saved.

Note that default unit of storage in Matlab is double which takes 64 bits. We do not

require that much precision, so we assign single precision uniformly which halves the

overall memory space. It also accelerates the calculation to a degree. More importantly,

this barely allows us to run the program on the Munich scenario with 16GB RAM. At

least 32GB RAM is recommended for source localization study on a larger scenario.

111

Figure 6-3: An example of Matlab profiling report when calculating propagation pathson the Arc de Triomphe scenario.

Matlab itself has a profiler which analyses processing time spent on functions and

syntaxes in full detail. An example of profiling the Arc de Triomphe scenario is shown

in Figure 6-3. The decomposition of self time allows us to trace a real bottleneck of

the program. Although ‘shoot ray10’ cost over 1190 seconds most of its processing

time was due to its sub-function called ‘MexFunction5448’ (see the next section for the

configuration of this CUDA function). As can be seen, this function took majority of

the total time and is being called for thousands of iterations which should be the focus

if better performance is expected. A parallel computing solution to address the issue

will be discussed in the next section.

Matlab provides powerful toolboxes on machine learning and data mining as well.

The Artificial Neural Network has been detailed in Chapter 4 to aid a navigation

solution. User customized ANN configurations, such as training algorithm, transfer

functions, layers and neurons etc. can be explored. In Chapter 5, Dynamic Time

Warping was applied to calculate fingerprint distances. A built-in DTW is available in

Matlab but could be further accelerated.

6.3 GPU-based acceleration

As can be seen from the profiling results in Figure 6-3, most of the time is spent

on hundreds of thousands of intersection detections. According to the literature [32],

Graphics Processing Units (GPU) are equipped with mathematical operation functions

in a parallel configuration which sees broader application in smart computation. It

112

has previously been noted that the intersection detections might benefit from GPU

parallel computing since they possess remarkable features of parallelism. A significant

characteristic during ray launching processing is that rays do not interfere along their

own propagation paths. This means from departure to arrival, rays are being calculated

independently which is advantageous for a GPU-based computing architecture. A

general schematic diagram of GPU can be found in Figure 6-4. The most significant

characteristic of this architecture is the large number of stream processors which operate

in parallel. Thus, the large number of ray interactions can be spread across these many

processors.

Hence, the intersection units has been reimplemented using Compute Unified Device

Architecture (CUDA) developed by NVIDIA. One advantage of CUDA is the existing

wide usage, and many libraries are available. It is an open architecture that allows

user customization and low-level manipulation. Last but not least, NVIDIA graphic

cards are readily available. The GPU used in this work is GeForce GTX 960M in a

Dell XPS15 laptop having a CUDA compute capability of 5.0 (Maxwell). This laptop

is equipped with an Intel i7 CPU with 16GB RAM.

Figure 6-4: A schematic diagram on GPU memory accessibility and multiprocessorconfiguration.

113

The intersection detection functions have been implemented in Visual Studio 2015

using the CUDA 8.0 toolkit. This time the intersection calculations are now assigned

in parallel to the GPU math units. Reading a ray and a surface coordinates from

the CPU memory, the GPU kernels ideally complete the intersection concurrently and

return the result back to host. Ideally, all ray-surface intersections can be dealt with

in a single loop.

In reality, however, one of the bottlenecks using GPU lies in the expensive commu-

nication between host and device. Memory allocation must be explicitly performed for

a CUDA thread to operate normally. The memory copy transfer rate between CPU and

GPU is relatively limited meaning that transferred data should be minimized where

possible. Moreover, threads are likely to be pipelined depending on the allocation of

threads per block and blocks per grid in terms of multiprocessor usage. As a result, in

common with all parallel computing approaches, the lowest throughput rate determines

the overall speed.

Table 6.1: An analysis of processing time with respect to number of walls and CPU orGPU based architecture

Configuration Time(s)

Number of surfaces Architecture Total Intersection Ratio(%)

S=25CPU 37.365 4.432 11.86

GPU 178.686 137.701 77.06

S=1752CPU 196.784 118.552 60.24

GPU 219.121 170.151 77.65

S=5728CPU 474.501 280.437 59.10

GPU 224.511 122.201 54.43

S=34890CPU 4452.890 3618.415 81.26

GPU 1094.517 396.038 36.18

After testing four scenarios of different complexity, processing time results using

CPU-based or GPU-based intersection detection when calculating one source are dis-

played in Table 6.1. It can be seen that when there are fewer walls, GPU-based perfor-

mance is rather poor (in fact, slower than CPU only), most likely due to expensive data

communication between host and GPU. As the number of walls increases, intersection

detection processing time goes up for both CPU or GPU. However, CPU-based inter-

section takes up a much larger proportion of the total time and becomes a significant

burden as can be more easily seen in Figure 6-5. GPU-based intersection units are

more capable of dealing with greater number of walls, which is typically the case in

114

real-world urban environments. In the Munich scenario with 34890 walls, the GPU-

based intersection detection function (which represents ≈36% of the total time) takes

one-seventh of the CPU processing time. Generally, GPU enables an overall three-fold

increase in computational performance in the simulation of typical urban environments.

Figure 6-5: Computation time (s) using CPU/GPU on scenarios of different complex-ity, where blue bar represents total processing time; red bar indicates time spent onthe intersection detection function; ’S’ is the number of surfaces presented in the envi-ronment.

The Visual Profiler, also developed by NVIDIA, is a popular tool in evaluating

CUDA performance. With the help of the Visual Profiler, Figure 6-6 shows a chart

of six iterations of the intersection detection unit using GPU. Obviously, most of the

time is spent on CUDA overheads in which the first-time initiation took up the great-

est proportion. Thankfully, if an intersection unit is called thousands of times it only

requires to set up once. A zoomed-in version of one iteration is displayed in Figure 6-7.

We notice that the computation in threads, marked with a blue bar, takes much less

than ‘cudaMalloc’ or ‘cudaMemcpy’ shown in brown. This implies the computation

bottleneck is indeed the communication between host and device. Thus, the accelera-

tion is yet far from completely optimized even though it is three times faster than pure

CPU-based architecture.

115

Figure 6-6: Profiling results on six iterations of the intersection function using theVisual Profiler.

Figure 6-7: Profiling results on one loop of the intersection function using the VisualProfiler.

A detailed GPU usage report can be found in Figure 6-8. The Compute/Memcpy =

0.418 indicates a less than 1% ratio between thread computing and host-device commu-

nication, which ideally needs to be larger for optimal computation. An ideal situation

is that threads are concentrating on local execution with little communication to the

host. This means that computation flow is in parallel fashion and fewer threads are

remaining idle for host-device communication. It is noted that GPU has vast potential

for further acceleration and could be part of future work.

Several efforts have been made to raise the GPU usage. The ‘Threads per Block’=1024

and ‘Blocks Dimensions’=[1024, 1024, 64] were allocated from trial and errors method

to determine the optimal configuration. These parameters determine how threads are

configured in the CUDA parallel architecture. Furthermore, a ‘cudaFree(0)’ instruction

is inserted before every iteration because it is the most swift procedure for waking-up

the GPU compared to directly starting memory allocation. This change alone shortens

the thread period by approximately 10%.

116

Figure 6-8: An analysis of GPU usage based on the intersection function developed.

As has been stated in [104], the local memory access is extremely fast within each

thread. However, if there is a race condition in accessing the same memory, delays will

occur which would greatly impact the performance. Since each ray is being searched

against all surfaces, regular assignment on the variables will result in surfaces competing

to request a ray for operation. One of the best solutions for this problem is the constant

memory which is globally visible across all threads of the GPU and easily accessible to

locals. There are only 64KB of constant memory in common GPUs offering extremely

high-speed reading, however, these cannot be written to. By placing the ray vector in

constant memory, the intersection detection function is typically accelerated by over

30%.

Finally, the GPU-based intersection detection outputs are delivered back to the

master program for subsequent calculations. This is implemented by Mex -function

encapsulation. The processing results of summoning CUDA as well as thread allocation

and execution are configured as a Mex -function to be returned to Matlab. Basically, the

outputs are comprised of flags, which tell the intersection type of a ray, and distances,

117

which indicate travelled distances starting from a virtual source to the next intersection

surface or edge. These are wrapped as a Mex array which is of the same format as

Matlab variables.

6.4 Summary

In this chapter, a sequence-based navigation scheme to enable off-line usage has been

proposed. This scheme can theoretically minimize the communication between user

and server during navigation. The feasibility of using Matlab optimization means and

GPU parallel computing techniques to accelerate our existing 2.5D ray launching model

have been investigated. Simplifications of Matlab built-in functions and lower precision

of variables was found to boost the processing speed. The use of parallel structures

embedded within the model has been analysed. A replacement CUDA intersection

detection function has been designed, tested and optimized. Compared to absolute

CPU computation, the GPU-based intersection detection function runs seven times

faster in large urban environments containing tens of thousands of walls.

Overall, it takes less than 190 hours to construct the off-line database of 616 source

candidates in the Munich scenario. The Matlab optimization and GPU-based acceler-

ation successfully multiply the computation speed of our ray launching model. These

optimizations make the proposed navigation and source localization systems more prac-

tical.

118

Chapter 7

Summary and future work

7.1 Summary

Navigation and radio source localization in difficult urban environments has drawn

significant industrial, military and civil engineering interest recently. However, the

facts are that traditional satellite and terrestrial navigation systems do not perform

well in complex urban environments. Currently, there is not a robust pervasive source

localization scheme. This work has investigated into the feasibility of utilizing a prop-

agation modelling approach, in which mature channel characterization capability has

been demonstrated, and then integrated into a mobile and radio source localization

solution. A fingerprinting technique is adopted to aid the development of the mobile

localization and source localization algorithms. Following data mining and pattern

recognition strategies, we are able to demonstrate reliable and accurate localization

algorithms which have been tested in different real-world scenarios.

In order to generate adequate channel characteristic ‘truth data’ for location train-

ing, we developed a dedicate 2.5D ray launching model suitable for urban propagation

simulation. The effectiveness of this model has been validated through checking against

field measurement results. For the Munich city center environment, an absolute mean

error of less than 8dB and a standard deviation of less than 7dB was found between

measurements and simulation results from thousands of RSS samples. The efficiency

of the model has been remarkably improved by optimizing the master program in

Matlab and by exploiting CUDA to handle the extremely time-consuming intersec-

tion detections. The OpenStreetMap project has allowed us to bring digital maps into

our environment modelling which requires precise shape and coordinate information of

buildings and trees.

The mobile localization algorithm has been shown to produce a mean error of

119

approximately 13m for a 700m×550m residential area with a maximum reflection

limit of seven. This algorithm is shown to remain robust against various errors and

perturbations. The location fingerprint is defined based on the simulated PDP and

angular spreads, which generate RSS, MED, RMSDS and AOAs parameters, with

respect to the mobile positions. The configuration of the ANN has been studied and

experimented with to yield an optimal trained network. A sequence-based localization

method has been proposed to minimize the communication between users and server

during navigation.

The source localization algorithm has been shown to confidently locate an emitter

to a pixel of 100m within a 2.4km×3.4km city center area even when exposed to a

significant degree of noise and different types of error. This is made possible with

the help of a UAV which is expected to collect low-resolution channel characteristic

parameters along a specified route. These parameters are then transformed into a

source location fingerprint which is searched against simulated fingerprints of source

candidates. Dynamic Time Warping has proved effective in calculating the fingerprint

similarities. The stability of this algorithm has been verified by corrupting the measured

fingerprint in different ways, such as the introduction of winds, reduction of sampling

rate, increase in AWGN level etc. We have also proposed a general route selection

strategy to increase the accuracy and reliability.

This thesis has broadly reviewed propagation modelling and localization methods

with a focus on their limitations. From this a novel urban navigation and source

localization solution has been developed and tested. New ideas and state-of-the-art

techniques have been applied to construct a comprehensive urban propagation model

and a new fingerprinted-based localization method. A key step in this work has been

the development of an efficient 2.5D ray launching engine, as well as exploitation of

recent advances in data mining and machine learning.

The accuracy and reliability of this localization system have been tested and con-

firmed. The author has also managed to optimize its performance for potential practical

applications. The body of this work has met the objectives and answered the questions

posed in the Chapter 1.

7.2 Future work

Although this thesis has answered the research questions posed in Chapter 1, many

new questions have arisen. Future work may include refinement and calibration of the

2.5D ray launching model, development of a 3D ray launching model, investigation

of the localization through deep learning on the digital maps directly, source local-

120

ization with unknown number of emitters and unknown mobility, and further GPU

accelerations.

Perhaps the most obvious area for further improvement is the 2.5D ray launch-

ing model. Currently, rays are always deemed as vertically polarized which is clearly

not true in real-world propagation. Comprehensive polarization of the electromagnetic

waves should be considered. Many cities are built on top of non-flat terrain which sig-

nificantly changes the propagation environment. Our model needs to take into account

the complex landscape, including the modelling of hills, valleys, lakes, cliffs etc. More-

over, further measurements and tests in different types of environments are required to

further calibrate the model and validate the localization algorithms.

Although it has been shown that 2.5D models are capable of channel modelling in

urban scenarios, 3D ray launching models, widely used in indoor propagation modelling,

allow even more accurate estimations. This is especially true for central business areas

where significant building height difference is seen. In these scenarios, the roof-top

diffraction model used in this thesis may not be valid. Based on our preliminary

3D modelling experiments in Chapter 3, we would suggest that a complete 3D ray

launching model is implemented to take advantage of more accelerations to reduce the

computational cost. In addition, the opportunistic sources are likely to be moving

in urban environments. This may drive the need for a 4D model, which regards the

channel characterization as a time-dependent function. A further option may also be

to track the source through the spatial Doppler frequencies generated.

In order to correlate channel characteristics with locations, we applied ANN to the

location fingerprints which were obtained by propagation modelling for one receiver

candidate at a time. A future attempt to shorten this pattern recognition phase might

be to directly train using the digital map and opportunistic sources to determine lo-

cation. The hypothesis here is that the ray launching step may not be required as

an intermediate step. In this case, the entire problem of fingerprinting to determine

location is passed to the ANN.

In the assumptions of the source localization algorithm, we only committed to locate

a single static emitter at a known frequency. It is worth investigating the more general

case of the localization of multiple emitters with different velocities. This would almost

certainly require a more powerful propagation modelling tool as well as more advanced

localization approaches.

As noted in Chapter 6, there are more tiers of parallelism that can be recognized

in the ray launching model and localization systems. For instance, rays shoot, prop-

agate and bounce independently; the reception detections for all reference points are

independent; the distance calculations between observed fingerprint and fingerprints in

121

the database are processed independently etc. These parallel structures are all options

for further GPU-based acceleration to lessen the computation time. A wider range of

acceleration schemes are likely to produce even better performance.

122

References

[1] MathWorks create, train, and simulate shallow and deep learning neural network.

https://uk.mathworks.com/products/neural-network.html. Accessed: 13

May 2017.

[2] OpenStreetMap official website. http://www.openstreetmap.org. Accessed: 30

September 2015.

[3] Overpass Turbo user interface. https://overpass-turbo.eu. Accessed: 30

September 2015.

[4] M. A. Abid and S. Cherkaoui. Wireless technology agnostic real-time localization

in urban areas. In 2011 IEEE 36th Conference on Local Computer Networks,

pages 727–733.

[5] S. Adelipour, M. Hamdollahzadeh, and F. Behnia. Constrained optimization of

sensors trajectories for moving source localization using TDOA and FDOA mea-

surements. In 2015 3rd RSI International Conference on Robotics and Mecha-

tronics (ICROM), pages 200–204.

[6] F. A. Agelet, F. P. Fontan, and A. Formella. Fast ray tracing for microcellular

and indoor environments. Magnetics, IEEE Transactions on, 33(2):1484–1487,

1997.

[7] F. A. Agelet, A. Formella, J. M. Hernando Rabanos, F. I. de Vicente, and F. P.

Fontan. Efficient ray-tracing acceleration techniques for radio propagation mod-

eling. Vehicular Technology, IEEE Transactions on, 49(6):2089–2104, 2000.

[8] F. Aguado, F. P. Fontan, and A. Formella. Indoor and outdoor channel simulator

based on ray tracing. In Vehicular Technology Conference, 1997, IEEE 47th,

volume 3, pages 2065–2069 vol.3.

123

https://uk.mathworks.com/products/neural-network.html

http://www.openstreetmap.org

https://overpass-turbo.eu

[9] T. Alwajeeh, P. Combeau, R. Vauzelle, and A. Bounceur. A high-speed 2.5D ray-

tracing propagation model for microcellular systems, application: Smart cities.

In EuCAP 2017.

[10] H. R. Anderson. A second generation 3-D ray-tracing model using rough surface

scattering. In Vehicular Technology Conference, 1996. Mobile Technology for the

Human Race., IEEE 46th, volume 1, pages 46–50 vol.1.

[11] H. R. Anderson. A ray-tracing propagation model for digital broadcast systems

in urban areas. Broadcasting, IEEE Transactions on, 39(3):309–317, 1993.

[12] L. Azpilicueta, M. Rawat, K. Rawat, F. M. Ghannouchi, and F. Falcone. A

ray launching neural network approach for radio wave propagation analysis in

complex indoor environments. Antennas and Propagation, IEEE Transactions

on, 62(5):2777–2786, 2014.

[13] K. Benzair. Measurements and modelling of propagation losses through vege-

tation at 1-4 GHz. In Antennas and Propagation, 1995., Ninth International

Conference on (Conf. Publ. No. 407), volume 2, pages 54–59 vol.2.

[14] I. Bisio, M. Cerruti, F. Lavagetto, M. Marchese, M. Pastorino, A. Randazzo, and

A. Sciarrone. A trainingless Wi-Fi fingerprint positioning approach over mobile

devices. IEEE Antennas and Wireless Propagation Letters, 13:832–835, 2014.

[15] D. Bojanjac, R. Nad, and Sisul G. Ray tracing model of pedestrian urban zone.

In ELMAR, 2010 PROCEEDINGS, pages 289–292.

[16] M. A. Caceres, F. Penna, H. Wymeersch, and R. Garello. Hybrid cooperative

positioning based on distributed belief propagation. IEEE Journal on Selected

Areas in Communications, 2011.

[17] K. A. Chamberlin and R. J. Luebbers. An evaluation of Longley-Rice and

GTD propagation models. Antennas and Propagation, IEEE Transactions on,

30(6):1093–1098, 1982.

[18] G. Chen, Y. Zhang, L. Xiao, J. Li, L. Zhou, and S. Zhou. Measurement-based

RSS-multipath neural network indoor positioning technique. In Electrical and

Computer Engineering (CCECE), 2014 IEEE 27th Canadian Conference on,

pages 1–7.

[19] C. Cheon, G. Liang, and H. L. Bertoni. Simulating radio channel statistics for

different building environments. IEEE Journal on Selected Areas in Communi-

cations, 19(11):2191–2200, 2001.

124

[20] S. Coco, A. Laudani, and L. Mazzurco. A novel 2-D ray tracing procedure for

the localization of EM field sources in urban environment. Magnetics, IEEE

Transactions on, 40(2):1132–1135, 2004.

[21] A. Coluccia and F. Ricciato. Maximum likelihood trajectory estimation of a

mobile node from RSS measurements. In 2012 9th Annual Conference on Wireless

On-Demand Network Systems and Services (WONS), pages 151–158.

[22] O. Corte, A. D. Gutierrez and J. M. Gomez. High-accuracy localization based on

the dominant rays of ray-tracing over fingerprinting techniques. In Antennas and

Propagation Society International Symposium (APSURSI), 2012 IEEE, pages 1–

2.

[23] M. Curry, B. Koala, M. Ciccotosto, and Y. Kuga. Indoor angle of arrival using

wide-band frequency diversity with experimental results and EM propagation

modeling. In Antennas and Propagation for Wireless Communications, 2000

IEEE-APS Conference on, pages 65–68.

[24] N. Czink, E. Bonek, Yin Xuefeng, and B. Fleury. Cluster angular spreads in a

MIMO indoor propagation environment. In Personal, Indoor and Mobile Radio

Communications, 2005. PIMRC 2005. IEEE 16th International Symposium on,

volume 1, pages 664–668.

[25] V. Degli-Esposti. Ray tracing propagation modelling: Future prospects. In

Antennas and Propagation (EuCAP), 2014 8th European Conference on, pages

2232–2232.

[26] H. Demuth and M. Beale. Neural Network Toolbox For Use with MATLAB. The

MathWorks, 2004.

[27] J. Deygout. Multiple knife-edge diffraction of microwaves. IEEE Transactions

on Antennas and Propagation, 14(4):480–489, 1966.

[28] G. Durgin, N. Patwari, and T. S. Rappaport. An advanced 3D ray launching

method for wireless propagation prediction. In Vehicular Technology Conference,

1997, IEEE 47th, volume 2, pages 785–789 vol.2.

[29] L. Felsen and N. Marcuvitz. Radiation and Scattering of Waves. The IEEE series

on Electromagnetic Wave Theory. John Wiley and Sons Inc., 1994.

[30] Y. Feng, L. Guo, P. Wang, and Z. Liu. Efficient ray-tracing model for propaga-

tion prediction for microcellular wireless communication systems. In Antennas,

125

Propagation and EM Theory (ISAPE), 2012 10th International Symposium on,

pages 432–435.

[31] C. Fernandez-Prades, L. L. Presti, and E. Falletti. Satellite radio localization

from GPS to GNSS and beyond: Novel technologies and applications for civil

mass market. Proceedings of the IEEE, (1882 - 1904), 2011.

[32] J. C. Fernndez-Gonzlez, M. L. Portugus, Y. A. Lpez, J. A. Lpez-Fernndez, and

F. A. Andrs. GPU-based computational acceleration of phaseless algorithms for

antenna characterization. In EuCAP 2017.

[33] V. Ferrara and C. M. Ottavi. Ray-tracing and UTD analyses for microcellular

communications in urban scene: a software methodology that uses 3-D GIS data

of buildings. In Remote Sensing and Data Fusion over Urban Areas, IEEE/ISPRS

Joint Workshop 2001, pages 255–259.

[34] C. Ghobadi, P. R. Shepherd, and S. R. Pennock. 2D ray-tracing model for in-

door radio propagation at millimetre frequencies, and the study of diversity tech-

niques. Microwaves, Antennas and Propagation, IEEE Proceedings, 145(4):349–

353, 1998.

[35] K. J. Gladstone and J. P. McGeehan. Computer simulation of multipath fading

in the land mobile radio environment. Electronic Circuits and Systems, IEEE

Proceedings G, 127(6):323–330, 1980.

[36] Mannesmann Mobilfunk GmbH. Cost 231 - urban micro cell measurements and

building data. 1999.

[37] N. Goddemeier, S. Rohde, and C. Wietfeld. Experimental validation of RSS

driven UAV mobility behaviors in IEEE 802.11s networks. In 2012 IEEE Globe-

com Workshops, pages 1550–1555.

[38] J. Goldhirsh and W. Vogel. Roadside tree attenuation measurements at UHF for

land mobile satellite systems. IEEE Transactions on Antennas and Propagation,

35(5):589–596, 1987.

[39] N. C. Goncalves and L. M. Correia. A propagation model for urban microcel-

lular systems at the UHF band. IEEE Transactions on Vehicular Technology,

49(4):1294–1302, 2000.

[40] E. Greenberg and Levy P. Propagation aspects for RF fingerprinting at open

areas over irregular terrain. In EuCAP 2017.

126

[41] Y. Gu, L. T. Hsu, and S. Kamijo. GNSS/onboard inertial sensor integration

with the aid of 3-D building map for lane-level vehicle self-localization in urban

canyon. IEEE Transactions on Vehicular Technology, 65(6):4274–4287, 2016.

[42] K. Guan, G. Li, D. He, et al. Spatial consistency of dominant components be-

tween ray-tracing and stochastic modeling in 3GPP high-speed train scenarios.

In EuCAP 2017.

[43] D. He, J. Yang, K. Guan, et al. Ray-tracing simulation and analysis of propaga-

tion for 3GPP high speed scenarios. In EuCAP 2017.

[44] J. Honda and T. Otsuyama. Rapid computation algorithm for radio wave prop-

agation characteristics on airport surface. In Complex, Intelligent and Software

Intensive Systems (CISIS), 2014 Eighth International Conference on, pages 302–

306.

[45] C. C. Huang and H. N. Manh. RSS-based indoor positioning based on multi-

dimensional kernel modeling and weighted average tracking. IEEE Sensors Jour-

nal, 16(9):3231–3245, 2016.

[46] S. Hussain and C. Brennan. An intra-visibility matrix based environment pre-

processing for efficient ray tracing. In EuCAP 2017.

[47] T. Imai and T. Fujii. Propagation loss in multiple diffraction using ray-tracing. In

Antennas and Propagation Society International Symposium, 1997. IEEE., 1997

Digest, volume 4, pages 2572–2575 vol.4.

[48] M. Jacob, S. Priebe, A. Maltsev, A. Lomayev, V. Erceg, and T. Kurner. A ray

tracing based stochastic human blockage model for the IEEE 802.11ad 60 GHz

channel model. In Antennas and Propagation (EuCAP), Proceedings of the 5th

European Conference on, pages 3084–3088.

[49] M. Jacob, S. Schon, U. Weinbach, and T. Kurner. Ray tracing supported pre-

cision evaluation for GPS indoor positioning. In Positioning, Navigation and

Communication, 2009. WPNC 2009. 6th Workshop on, pages 15–22.

[50] S. Jiang and S. V. Georgakopoulos. Optimum wireless power transmission from

air to lossy media. In Wireless and Microwave Technology Conference (WAMI-

CON), 2011 IEEE 12th Annual, pages 1–6.

[51] Y. Jin, N. O’Donoughue, and J. M. F. Moura. Position location by time reversal in

communication networks. In 2008 IEEE International Conference on Acoustics,

Speech and Signal Processing, pages 3001–3004.

127

[52] J. Karedal, F. Tufvesson, T. Abbas, O. Klemp, A. Paier, L. Bernado, and A. F.

Molisch. Radio channel measurements at street intersections for vehicle-to-vehicle

safety applications. In Vehicular Technology Conference (VTC 2010-Spring),

2010 IEEE 71st, pages 1–5.

[53] A. O. Kaya, L. Greenstein, D. Chizhik, R. Valenzuela, and N. Moayeri. Emitter

localization and visualization (ELVIS): A backward ray tracing algorithm for

locating emitters. In Information Sciences and Systems, 2007. CISS ’07. 41st

Annual Conference on, pages 376–381.

[54] Joseph B. Keller. Geometrical theory of diffraction. Journal of the Optical Society

of America, 52(2):116–130, 1962.

[55] J. M. Kelner, C. Ziolkowski, L. Nowosielski, and M. Wnuk. Localization of

emission source in urban environment based on the Doppler effect. In 2016 IEEE

83rd Vehicular Technology Conference (VTC Spring), pages 1–5.

[56] E. Keogh and Ratanamahatana C. A. Exact indexing of dynamic time warping.

Knowledge and Information Systems, 2004.

[57] Y. Jun Y. Chong Kim, J. H. Yoon and H. J. Hong. Millimeter-wave delay spread

measurement and simulation at LoS urban low-rise environments. In Informa-

tion and Communication Technology Convergence (ICTC), 2015 International

Conference on, pages 1194–1196.

[58] C. Kleist. Time Series Data Mining Methods: A Review. Master’s thesis, School

of Business and Economics, Humboldt University of Berlin, 2015.

[59] T. Kopacz, S. Schiebl, and D. Heberling. Investigating the overestimation of

base station exposure in urban environments due to assumption of free space

propagation. In EuCAP 2017.

[60] R. G. Kouyoumjian and P. H. Pathak. A uniform geometrical theory of diffrac-

tion for an edge in a perfectly conducting surface. Proceedings of the IEEE,

62(11):1448–1461, 1974.

[61] Y. Kurdi and L. de Haro. DOA measurements on indoor channel based on mu-

sic and MDL processing. In Antennas and Propagation Society International

Symposium, 2008. AP-S 2008. IEEE, pages 1–4.

[62] T. Kurner, D. J. Cichon, and W. Wiesbeck. Concepts and results for 3D digital

terrain-based wave propagation models: an overview. IEEE Journal on Selected

Areas in Communications, 11(7):1002–1012, 1993.

128

[63] M. Laaraiedh, N. Amiot, and B. Uguen. Efficient ray tracing tool for UWB

propagation and localization modeling. In Antennas and Propagation (EuCAP),

2013 7th European Conference on, pages 2307–2311.

[64] Z. Lai. The Development of an Intelligent Ray Launching Algorithm for Wireless

Network Planning. PhD thesis, University of Bedfordshire, 2010.

[65] Z. Lai, H. Song, P. Wang, H. Mu, L. Wu, and J. Zhang. Implementation and

validation of a 2.5D intelligent ray launching algorithm for large urban scenarios.

In 2012 6th European Conference on Antennas and Propagation (EuCAP), pages

2396–2400.

[66] M. C. Lawton and J. P. McGeehan. The application of GTD and ray launching

techniques to channel modelling for cordless radio systems. In Vehicular Tech-

nology Conference, 1992, IEEE 42nd, pages 125–130 vol.1.

[67] J. F. Li, J. Wagen and E. Lachat. Propagation over rooftop and in the horizontal

plane for small and micro-cell coverage predictions. In Vehicular Technology

Conference, 1997, IEEE 47th, volume 2, pages 1123–1127 vol.2.

[68] X. Li. Efficient Ray Tracing Simulation. Master’s thesis, Lund University and

Forschungszentrum Telekommunikation Wien GmbH, 2014.

[69] Y. Ding Y. Eghbali H. Muhaidat S. H. Li, L. Ibdah and X. Ma. Indoor multi-wall

path loss model at 1.93 GHz. In Military Communications Conference, MILCOM

2013 - 2013 IEEE, pages 1233–1237.

[70] G. Liang and H. L. Bertoni. A new approach to 3D ray tracing for site specific

propagation modeling. In Vehicular Technology Conference, 1997, IEEE 47th,

volume 2, pages 1113–1117 vol.2.

[71] G. Liang and H. L. Bertoni. A new approach to 3-D ray tracing for propagation

prediction in cities. IEEE Transactions on Antennas and Propagation, 46(6):853–

863, 1998.

[72] J. Lin and Y. Li. Finding structural similarity in time series data using BOP. In-

ternational Conference on Scientific and Statistical Database Management, pages

461–477, 2009.

[73] J. Lin, S. Williamson, K. Borne, and D. DeBarr. Pattern recognition in time

series. Advances in Machine Learning and Data Mining for Astronomy, pages

617–645, 2010.

129

[74] Z. Liu, L. Guo, X. Meng, and Z. Zhong. A novel 3D ray-tracing model for

propagation prediction in indoor environments. In Antennas, Propagation and

EM Theory (ISAPE), 2012 10th International Symposium on, pages 428–431.

[75] Z. Y. Liu, L. X. Guo, and C. L. Li. Effects of antenna polarization on power and

RMS delay spread in LOS/OOS indoor radio channel. In General Assembly and

Scientific Symposium (URSI GASS), 2014 XXXIth URSI, pages 1–4.

[76] D. C. Livingston. The physics of microwave propagation. Prentice-Hall electrical

engineering series: Microwaves and fields series. Prentice-Hall, 1970.

[77] A. Lodhi. Prediction study of LTE received signal strengths at short distances

from base stations. Delivery report, SIRADEL Ltd and OFCOM Ltd, 2013.

[78] H. Lohrasbipeydeh, T. A. Gulliver, and H. Amindavar. Unknown transmit power

RSSD based source localization with sensor position uncertainty. IEEE Transac-

tions on Communications, 63(5):1784–1797, 2015.

[79] K. W. K. Lui and H. C. So. Range-based source localisation with pure reflector

in presence of multipath propagation. Electronics Letters, 46(13):957–958, 2010.

[80] L. R. Maciel, H. L. Bertoni, and H. N. Xia. Unified approach to prediction of

propagation over buildings for all ranges of base station antenna height. IEEE

Transactions on Vehicular Technology, 42(1):41–45, 1993.

[81] B. Malila, O. Falowo, and N. Ventura. Millimeter wave small cell backhaul: An

analysis of diffraction loss in NLOS links in urban canyons. In AFRICON, 2015,

pages 1–5.

[82] S. Matsuo, M. Gocho, T. Hashimoto, and A. Ozaki. GPU-based parallel algorithm

for VPL-approximated EM wave propagation. In EuCAP 2017.

[83] P. Meissner, M. Gan, F. Mani, E. Leitinger, M. Frohle, C. Oestges, T. Zemen,

and K. Witrisal. On the use of ray tracing for performance prediction of UWB

indoor localization systems. In Communications Workshops (ICC), 2013 IEEE

International Conference on, pages 68–73.

[84] B. Meng, W. Shan, H. Li, and F. Li. Construction and implement of semi-physical

electromagnetic propagation simulation platform. In Electronics, Communica-

tions and Control (ICECC), 2011 International Conference on, pages 783–787.

[85] Y. Miao, Q. Gueuning, M. Gan, and C. Oestges. Adding diffuse scattering cor-

relation to effective roughness models in ray tracing. In EuCAP 2017.

130

[86] Harold Mott. Polarization in Antennas and Radar. Wiley-Interscience publica-

tion, 1986.

[87] J. Nuckelt, T. Abbas, F. Tufvesson, C. Mecklenbrauker, L. Bernado, and

T. Kurner. Comparison of ray tracing and channel-sounder measurements for

vehicular communications. In Vehicular Technology Conference (VTC Spring),

2013 IEEE 77th, pages 1–5.

[88] J. Nuckelt, D. M. Rose, T. Jansen, and T. Kurner. On the use of openstreetmap

data for V2X channel modeling in urban scenarios. In Antennas and Propagation

(EuCAP), 2013 7th European Conference on, pages 3984–3988.

[89] P. Okvist, H. Asplund, A. Simonsson, B. Halvarsson, J. Medbo, and N. Seifi. 15

GHz propagation properties assessed with 5G radio access prototype. In Personal,

Indoor, and Mobile Radio Communications (PIMRC), 2015 IEEE 26th Annual

International Symposium on, pages 2220–2224.

[90] D. Pena, R. Feick, H. D. Hristov, and W. Grote. Measurement and modeling of

propagation losses in brick and concrete walls for the 900-MHz band. Antennas

and Propagation, IEEE Transactions on, 51(1):31–39, 2003.

[91] H. Peter. Machine Learning In Action. Manning Publications Co., Shelter Island,

NY 11964, 2012.

[92] B. R. Phelan, E. H. Lenzing, and R. M. Narayanan. Source localization using

unique characterizations of multipath propagation in an urban environment. In

Sensor Array and Multichannel Signal Processing Workshop (SAM), 2012 IEEE

7th, pages 189–192.

[93] Y. Pinhasi, A. Yahalom, and S. Petnev. Propagation of ultra wide-band signals

in lossy dispersive media. In Microwaves, Communications, Antennas and Elec-

tronic Systems, 2008. COMCAS 2008. IEEE International Conference on, pages

1–10.

[94] L. A. R. Ramirez, L. A. R. Silva Mello, and F. J. V. Hasselmannn. Application of

a ray tracing technique for coverage predictions in micro cellular environments.

In Microwave and Optoelectronics Conference, 2003. IMOC 2003. Proceedings of

the 2003 SBMO/IEEE MTT-S International, volume 1, pages 171–175 vol.1.

[95] J. Richter, M. O. Al-Nuaimi, and L. P. Ivrissimtzis. Optimisation of radio cov-

erage in urban microcells using a UTD based ray-tracing model. Microwaves,

Antennas and Propagation, IEEE Proceedings, 151(3):187–192, 2004.

131

[96] J. Richter, R. F. S. Caldeirinha, M. O. Al-Nuaimi, A. Seville, N. C. Rogers,

and N. Savage. A generic narrowband model for radiowave propagation through

vegetation. In 2005 IEEE 61st Vehicular Technology Conference, volume 1, pages

39–43 Vol. 1.

[97] K. Rizk, A. Mawira, J. F. Wagen, and F. Gardiol. Propagation in urban micro-

cells with high rise buildings. In Vehicular Technology Conference, 1996. Mobile

Technology for the Human Race., IEEE 46th, volume 2, pages 859–863 vol.2.

[98] N. C. Rogers. A generic model of 1-60 GHz radio propagation through vegetation.

Report, Radiocommunications Agency, 2002.

[99] J. A. Romo, I. F. Anitzine, and F. P. Fontan. Application of neural networks to

field strength prediction for indoor environments. In Antennas and Propagation,

2006. EuCAP 2006. First European Conference on, pages 1–6.

[100] C. Saeidi, F. Hodjatkashani, and A. Fard. New tube-based shooting and bouncing

ray tracing method. In Advanced Technologies for Communications, 2009. ATC

’09. International Conference on, pages 269–273.

[101] A. A. M. Saleh and R. A. Valenzuela. A statistical model for indoor multipath

propagation. Selected Areas in Communications, IEEE Journal on, 5(2):128–137,

1987.

[102] S. Salous. Radio Propagation Measurement and Channel Modelling. Wiley, 1st

edition, 2013.

[103] S. Salous, A. Cheema, and X. Raimundo. Radio channel propagation measure-

ments using a multiband agile chirp sounder. In General Assembly and Scientific

Symposium (URSI GASS), 2014 XXXIth URSI, pages 1–4.

[104] J. Sanders and K. Edward. CUDA by Example: an introduction to general-

purpose GPU programming. Pearson Education Inc., 2010.

[105] S. Saunders. Antennas and Propagation for Wireless Communication Systems.

John Wiley and Sons Ltd, 2nd edition, 2007.

[106] K. R. Schaubach, N. J. Iv Davis, and T. S. Rappaport. A ray tracing method for

predicting path loss and delay spread in microcellular environments. In Vehicular

Technology Conference, 1992, IEEE 42nd, pages 932–935 vol.2.

[107] M. Schiller, A. Knoll, M. Mocker, and T. Eibert. GPU accelerated ray launching

for high-fidelity virtual test drives of VANET applications. In High Performance

132

Computing and Simulation (HPCS), 2015 International Conference on, pages

262–268.

[108] R. Schmidt. Multiple emitter location and signal parameter estimation. IEEE

Transactions on Antennas and Propagation, 34(3):276–280, 1986.

[109] A. Schmitz and L. Kobbelt. Efficient and accurate urban outdoor radio wave

propagation. In Electromagnetics in Advanced Applications (ICEAA), 2011 In-

ternational Conference on, pages 323–326.

[110] J. Schmitz, F. Schrder, and R. Mathar. TDOA fingerprinting for localization in

non-line-of-sight and multipath environments. In 2015 International Symposium

on Antennas and Propagation (ISAP), pages 1–4.

[111] E. Semaan, F. Harrysson, A. Furuskar, and H. Asplund. Outdoor-to-indoor

coverage in high frequency bands. In Globecom Workshops (GC Wkshps), 2014,

pages 393–398.

[112] S. A. A. Shahidian and H. Soltanizadeh. Autonomous trajectory control for

limited number of aerial platforms in RF source localization. In Robotics and

Mechatronics (ICROM), 2015 3rd RSI International Conference on, pages 755–

760.

[113] Y. Shu, Y. Huang, J. Zhang, P. Coue, P. Cheng, J. Chen, and K. G. Shin.

Gradient-based fingerprinting for indoor localization and tracking. IEEE Trans-

actions on Industrial Electronics, 63(4):2424–2433, 2016.

[114] T. Sofos and P. Constantinou. Propagation model for vegetation effects in ter-

restrial and satellite mobile systems. IEEE Transactions on Antennas and Prop-

agation, 52(7):1917–1920, 2004.

[115] S. P. Sotiroudis, S. K. Goudos, K. A. Gotsis, K. Siakavara, and J. N. Saha-

los. Application of a composite differential evolution algorithm in optimal neural

network design for propagation path-loss prediction in mobile communication

systems. Antennas and Wireless Propagation Letters, IEEE, 12:364–367, 2013.

[116] Q. Spencer, M. Rice, B. Jeffs, and M. Jensen. A statistical model for angle of

arrival in indoor multipath propagation. In Vehicular Technology Conference,

1997, IEEE 47th, volume 3, pages 1415–1419 vol.3.

[117] S. Spiegel, J. Gaebler, A. Lommatzsch, E. D. Luca, and S. Albayrak. Pattern

recognition and classification for multivariate time series. SensorKDD’11, 2011.

133

[118] A. Tayebi, J. Gomez, F. S. de Adana, and O. Gutierrez. Ray-tracing application

to mobile localization in multipath indoor environments. In Electromagnetics in

Advanced Applications, 2009. ICEAA ’09. International Conference on, pages

412–415.

[119] F. Tila, P. R. Shepherd, and S. R. Pennock. Indoor ray tracing evaluation of en-

hanced high frequency communications using directional antennas at the bases-

tation and space diversity at the terminal. In Microwave Conference, 2001. 31st

European, pages 1–5.

[120] S. A. Torrico, K. L. Chee, and T. Kurner. A propagation prediction model in veg-

etated residential environments - a simplified analytical approach. In Proceedings

of the 5th European Conference on Antennas and Propagation (EuCAP), pages

3279–3283.

[121] C. Tzaras and S. R. Saunders. Rapid, uniform computation of multiple knife-edge

diffraction. Electronics Letters, 35(3):237–239, 1999.

[122] F. T. Ulaby, R. K. Moore, and A. K. Fung. Microwave Remote Sensing Fun-

damentals and Radiometry, volume 1 of Microwave Remote Sensing Active and

Passive. ARTECH HOUSE, 1981.

[123] R. L. Vacarro and J. J. Fante. Cancellation of jammers and jammer multipath in

a GPS. IEEE Aerospace and Electronic Systems Magazine, pages 25 – 28, 1998.

[124] N. Sipus Z. Vilovic, I. Burum. A comparison of neural network models for indoor

field strength prediction. In ELMAR, 2007, pages 251–254.

[125] W. J. Vogel and G. W. Torrence. Propagation measurements for satellite radio

reception inside buildings. Antennas and Propagation, IEEE Transactions on,

41(7):954–961, 1993.

[126] A. Wadhwa, U. Madhow, J. Hespanha, and B. M. Sadler. Following an RF trail to

its source. In 2011 49th Annual Allerton Conference on Communication, Control,

and Computing (Allerton), pages 580–587.

[127] B. Wagner, D. Timmermann, G. Ruscher, and T. Kirste. Device-free user local-

ization utilizing artificial neural networks and passive RFID. In 2012 Ubiquitous

Positioning, Indoor Navigation, and Location Based Service (UPINLBS), pages

1–7.

134

[128] D. J. Walter, K. Bryan, J. Stephens, C. Bullmaster, and V. Chakravarthy. Lo-

calization of RF emitters using compressed sensing with multiple cooperative

sensors. In 2012 IEEE National Aerospace and Electronics Conference (NAE-

CON), pages 236–240.

[129] P. Wawrzyniak, S. Hausman, and P. Korbel. Area based indoor tracking algo-

rithm based on sequence detection and maximum likelihood metrics. In 2016

10th European Conference on Antennas and Propagation (EuCAP), pages 1–3.

[130] P. Wawrzyniak, S. Hausman, and P. Korbel. Sequence detection of movement

for accurate area based indoor positioning and tracking. In 2015 9th European

Conference on Antennas and Propagation (EuCAP), pages 1–4.

[131] J. Weng, X. Tu, Z. Lai, S. Salous, and J. Zhang. Modelling the mm-wave channel

based on intelligent ray launching model. In Antennas and Propagation (EuCAP),

2015 9th European Conference on, pages 1–4.

[132] R. Wilfinger, T. Moder, M. Wieser, and B. Grosswindhager. Indoor position

determination using location fingerprinting and vehicle sensor data. In 2016

European Navigation Conference (ENC), pages 1–9.

[133] X. Wu, V. Kumar, J. Ross, et al. Top 10 algorithms in data mining. In IEEE

International Conference on Data Mining (ICDM), 2007.

[134] H. H. Xia. An analytical model for predicting path loss in urban and suburban

environments. In Personal, Indoor and Mobile Radio Communications, 1996.

PIMRC’96., Seventh IEEE International Symposium on, volume 1, pages 19–23

vol.1.

[135] C. Yang, B. Wu, and C. Ko. A ray-tracing method for modeling indoor wave

propagation and penetration. Antennas and Propagation, IEEE Transactions on,

46(6):907–919, 1998.

[136] G. Yang, K. Pahlavan, and J. F. Lee. A 3D propagation model with polar-

ization characteristics in indoor radio channels. In Global Telecommunications

Conference, 1993, including a Communications Theory Mini-Conference. Tech-

nical Program Conference Record, IEEE in Houston. GLOBECOM ’93., IEEE,

pages 1252–1256 vol.2.

[137] G. Yang, K. Pahlavan, J. F. Lee, A. J. Dagen, and J. Vancraeynest. Prediction of

radio wave propagation in four blocks of New York city using 3D ray tracing. In

Personal, Indoor and Mobile Radio Communications, 1994. Wireless Networks -

135

Catching the Mobile Future., 5th IEEE International Symposium on, volume 1,

pages 263–267 vol.1.

[138] S. Yun-di Lien and M. Cherniakov. Analytical approach for multipath delay

spread power distribution. In Global Telecommunications Conference, 1998.

GLOBECOM 1998. The Bridge to Global Integration. IEEE, volume 6, pages

3680–3685 vol.6.

[139] H. Zhao, G. Zheng, L. Zhang, W. Fu, H. Wei, and H. Shao. Ray tracing analysis of

Gaussian beamed millimetre wave propagation in circle tunnel. In Wireless Mobile

and Computing (CCWMC 2009), IET International Communication Conference

on, pages 164–167.

[140] Y. Zhao and Y. Zhuo. The algorithm of position related channel character extrac-

tion and match localization based on ray-tracing. In Advanced Computer Theory

and Engineering (ICACTE), 2010 3rd International Conference on, volume 2,

pages V2–429–V2–433.

[141] H. Zou, B. Huang, X. Lu, H. Jiang, and L. Xie. A robust indoor positioning

system based on the procrustes analysis and weighted extreme learning machine.

IEEE Transactions on Wireless Communications, 15(2):1252–1266, 2016.

136

Appendix

A. Overpass Turbo map geometry extraction function based on Matlab

R2016b

Function S = MapExtract3( ) % Convert Overpass Turbo coordinates into Matlab

structure files

Place0 = kml2struct(’Hoxton1.kml’); % Read structural data of the Hoxton scenario

from file

Os = [ -0.093430, 51.528708]; % The origin for Hoxton scenario

rmax = 6378137; % Maximum radius of the earth

rmin = 6356752.3142; % Minimum radius of the earth

r = sqrt(((rmax2· cos(Os(2)))2 + (rmin2· sin(Os(2)))2)/

((rmax·cos(Os(2)))2 + (rmin·sin(Os(2)))2));

i=1;

for t=1:size(Place0,2)

if strcmp(Bath0(t).Name, ’undefined’)==0

Place1,i = [Place0(t).Lon Place0(t).Lat];

i=i+1;

end;

end;

Place S = cell(1,i-1);

for q=1:size(Place,2)

P = size(Place{1,q},1);

for p=1:P

if isnan(Place{1,q}(p,:)) =[0 0]

break;

else

Place S{1,q}(p,:) = [sind(Place{1,q}(p,1)-Os(1))·r, sind(Place{1,q}(p,2)-Os(2))·r];end;

end;

end;

S = [ ];

for q=1:size(Place S,2)

137

P = size(Place S{1,q},1);

for p=1:(P-1)

S(size(S,1)+1,:) = PlaceS {1,q }(p,:);

S(size(S,1)+1,:) = PlaceS {1,q }(p+1,:);

end;

end;

m = length(S)/2;

figure(33);

hold on;

grid on;

for k=1:m

a = min(S(2*k-1,1),S(2*k,1));

b = max(S(2*k-1,1),S(2*k,1));

if (a =b)

x = a:0.01:b;

y = ((S(2*k,2)-S(2*k-1,2))/(S(2*k,1)-S(2*k-1,1)))*(x-S(2*k-1,1))+S(2*k-1,2);

plot(x,y,’k-’,’linewidth’,2);

else

a0 = min(S(2*k-1,2),S(2*k,2));

b0 = max(S(2*k-1,2),S(2*k,2));

y = a0:0.01:b0;

plot(a*ones(size(y)),y,’k-’,’linewidth’,2);

end;

end;

axis equal;

end

138

B. Ray-surface intersection detection based on Visual Studio Commu-

nity 2015 using CUDA8.0

# include ”stdafx.h”

# include <stdio.h>

# include <stdlib.h>

# include <iostream>

# include <math.h>

# include <device launch parameters.h>

# include <cuda runtime.h>

# include <cuda.h>

# include ”mex.h”

# define n 2

# define S 5448

# define Pi 3.1415926535897932384626

void intersect0(float *dev a , float *dev b , float *dev t , int *dev flag, float *Ori,

float *Dir);

void mexFunction(int nlhs, mxArray *plhs[ ], int nrhs, const mxArray *prhs[ ]) {

nrhs = 4;

double *valueO = mxGetPr(prhs[0]);

double *valueD = mxGetPr(prhs[1]);

float origin[n], direction[n];

origin[0] = (float)valueO[0];

origin[1] = (float)valueO[1];

direction[0] = (float)valueD[0];

direction[1] = (float)valueD[1];

mwSize row, col;

mwSize nrow = mxGetM(prhs[2]);

mwSize ncol = mxGetN(prhs[2]);

double *values = mxGetPr(prhs[2]);

double *valuesB = mxGetPr(prhs[3]);

cudaFree(0);

int *dev flag, flag[S];

float *dev t, t[S], *dev a, *dev b;

139

float a[S][n];

float b[S][n];

cudaHostAlloc((void**)& a, n*S * sizeof(float), cudaHostAllocDefault);

cudaHostAlloc((void**)& b, n*S * sizeof(float), cudaHostAllocDefault);

for (row = 0; row < nrow; row++) {for (col = 0; col < ncol; col++) {a[row][col] = (float)values[nrow * col + row];

b[row][col] = (float)valuesB[nrow * col + row];

}}cudaHostAlloc((void**)& flag, S * sizeof(int), cudaHostAllocDefault);

cudaHostAlloc((void**)& t, S * sizeof(float), cudaHostAllocDefault);

cudaMalloc((void**)& dev flag, S * sizeof(int));

cudaMalloc((void**)& dev t, S * sizeof(float));

cudaMalloc((void**)& dev a, n*S * sizeof(float));

cudaMalloc((void**)& dev b, n*S * sizeof(float));

cudaMemcpyAsync(dev a, a, n*S * sizeof(float), cudaMemcpyHostToDevice);

cudaMemcpyAsync(dev b, b, n*S * sizeof(float), cudaMemcpyHostToDevice);

intersect0(dev a, dev b, dev t, dev flag, origin, direction);

cudaMemcpyAsync(flag, dev flag, S * sizeof(int), cudaMemcpyDeviceToHost);

cudaMemcpyAsync(t, dev t, S * sizeof(float), cudaMemcpyDeviceToHost);

nlhs = 2;

double mexFlag[S], mexT[S];

for (int j = 0; j < S; j++) {mexFlag[j] = (double)flag[j];

mexT[j] = (double)t[j];

}plhs[0] = mxCreateDoubleMatrix(1, S, mxREAL);

plhs[1] = mxCreateDoubleMatrix(1, S, mxREAL);

memcpy(mxGetPr(plhs[0]), mexFlag, S * sizeof(double));

memcpy(mxGetPr(plhs[1]), mexT, S * sizeof(double));

cudaFreeHost(a);

cudaFreeHost(b);

cudaFree(dev a);

140

cudaFree(dev b);

cudaFreeHost(t);

cudaFreeHost(flag);

cudaFree(dev t);

cudaFree(dev flag);

return;

}

constant float o[n];

constant float d[n];

global void intersectMEX(float *a, float *b, float *t, int *flag) {

int id = blockIdx.x * blockDim.x + threadIdx.x;

if (id >= S) {return;

}if (id < S) {float line[n] = { a[id*n + 0] - b[id*n + 0], a[id*n + 1] - b[id*n + 1] };float l[n] = { line[0] / sqrt(line[0] * line[0] + line[1] * line[1]), line[1] / sqrt(line[0] *

line[0] + line[1] * line[1]) };float ParallelCheck = -l[1] * d[0] + l[0] * d[1];

float x0;

if (d[0] < 0) {x0 = -d[0];

}else {x0 = d[0];

}float y0;

if (d[1] < 0) {

141

y0 = -d[1];

}else {y0 = d[1];

}if (ParallelCheck <= 0.0001 && ParallelCheck >= -0.0001) {float ao[ ] = { a[id*n + 0] - o[0], a[id*n + 1] - o[1] };float bo[ ] = { b[id*n + 0] - o[0], b[id*n + 1] - o[1] };float OverlapCheck = -ao[1] * bo[0] + ao[0] * bo[1];

if (OverlapCheck <= 0.0001 && OverlapCheck >= -0.0001) {if (x0 < 0.01) {float ta = a[id*n + 1] - o[1];

if (ta < 0) { ta = -ta; }float tb = b[id*n + 1] - o[1];

if (tb < 0) { tb = -tb; }float tAB[2][n] = { ao[0],ao[1],bo[0],bo[1] };float tt[2] = { ta, tb };if ((tt[0] + tt[1]) > sqrt(line[0] * line[0] + line[1] * line[1])) {int index;

if (tt[0] < tt[1]) {t[id] = tt[0];

index = 0;

}else {t[id] = tt[1];

index = 1;

}if ((tAB[index][0] * d[0] + tAB[index][1] * d[1]) > 0) {flag[id] = -1;

return;

}else {flag[id] = 0;

t[id] = 0;

return;

}}

142

else {flag[id] = 0;

t[id] = 0;

return;

}}if (y0 < 0.01) {float ta = a[id*n + 0] - o[0];

if (ta < 0) { ta = -ta; }float tb = b[id*n + 0] - o[0];

if (tb < 0) { tb = -tb; }float tAB[2][n] = { ao[0],ao[1],bo[0],bo[1] };float tt[2] = { ta, tb };if ((tt[0] + tt[1]) > sqrt(line[0] * line[0] + line[1] * line[1])) {int index;

if (tt[0] < tt[1]) {t[id] = tt[0];

index = 0;


index = 1;


return;


t[id] = 0;

return;

}}else {flag[id] = 0;

t[id] = 0;

return;

143

}}float ta = (a[id*n + 1] - o[1]) / d[1];

if (ta < 0) { ta = -ta; }float tb = (b[id*n + 1] - o[1]) / d[1];

if (tb < 0) { tb = -tb; }float tAB[2][n] = { ao[0],ao[1],bo[0],bo[1] };float tt[2] = { ta, tb };if ((tt[0] + tt[1]) > (0.1 + sqrt(line[0] * line[0] + line[1] * line[1]))) {int index;

if (tt[0] < tt[1]) {t[id] = tt[0];

index = 0;


index = 1;


return;


t[id] = 0;

return;

}}}flag[id] = 0;

t[id] = 0;

return;

}float k;

if (d[0]<0.01 && d[0]>-0.01) {k = (o[0] - a[id*n + 0]) / (b[id*n + 0] - a[id*n + 0]);

t[id] = (a[id*n + 1] - o[1] + k*(b[id*n + 1] - a[id*n + 1])) / d[1];

144

}else {float nomi = (a[id*n + 1] - o[1])*d[0] - (a[id*n + 0] - o[0])*d[1];

float deno = (b[id*n + 0] - a[id*n + 0])*d[1] - (b[id*n + 1] - a[id*n + 1])*d[0];

k = nomi / deno;

t[id] = ((a[id*n + 0] - o[0]) + k*(b[id*n + 0] - a[id*n + 0])) / d[0];

}if (t[id] < 0.1) {flag[id] = 0;

t[id] = 0;

return;

}else if (k <= -0.001 || k >= 1.001) {flag[id] = 0;

t[id] = 0;

return;

}else {if (k > -0.001 && k < 0.001) {flag[id] = -2;

return;

}else if (k > 0.999 && k < 1.001) {flag[id] = -2;

return;


if (d[0]<0.01 && d[0]>-0.01) { t[id] − = 0.1; }}}}

}

145

void intersect0(float *dev a, float *dev b, float *dev t, int *dev flag, float *origin,

float *direction) {

cudaMemcpyToSymbol(o, origin, n * sizeof(float));

cudaMemcpyToSymbol(d, direction, n * sizeof(float));

dim3 blocksPerGrid((S + 511) / 512, 1);

dim3 threadsPerBlock(512, 1)

intersectMEX <<< blocksPerGrid, threadsPerBlock >>> (dev a, dev b, dev t,

dev flag);

}

146

C. List of publications

[1]. Z. Dai and R. J. Watson, “A GPU accelerated ray-launching model using

CUDA,” 2018 12th European Conference on Antennas and Propagation (EuCAP),

London, 2018.

[2]. Z. Dai, P. R. Shepherd and R. J. Watson, “UAV-aided source localization in

urban environments based on ray launching simulation,” 2017 IEEE AP-S Symposium

on Antennas and Propagation and USNC-URSI Radio Science Meeting, San Diego,

USA, 2017.

[3]. Z. Dai, R. J. Watson and P. R. Shepherd, “A propagation modeling approach to

navigation,” 2017 11th European Conference on Antennas and Propagation (EuCAP),

Paris, 2017.

[4]. Z. Dai, R. J. Watson and P. R. Shepherd, “A propagation modeling approach

to source location and navigation,” 2016 10th European Conference on Antennas and

Propagation (EuCAP), Davos, 2016.

147

Documents

Propagation Modelling for Urban Source …...experience which always guide me towards a right direction, and o ering me an open minded research environment. His dedication and commitment